doi
stringlengths 10
10
| chunk-id
stringlengths 1
4
| chunk
stringlengths 1
1.66k
| id
stringlengths 10
10
| title
stringlengths 19
148
| summary
stringlengths 345
1.92k
| source
stringlengths 31
31
| authors
list | categories
list | comment
stringlengths 4
284
⌀ | journal_ref
stringclasses 14
values | primary_category
stringclasses 16
values | published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1811.03600
|
79
|
Batch Size0.060.080.100.120.140.160.1810 epochs
20 epochs
50 epochs
100 epochs200 epochs
500 epochs
1000 epochs
10000 epochs
(b)Simple CNN on Fashion MNIST
242526272829210211212213214215
Batch Size4.04.24.44.64.85.05.25.4Best Validation Cross Entropy10k steps
20k steps
50k steps
100k steps
200k steps
500k steps
1000k steps
242526272829210211212213214215
Batch Size4.04.24.44.64.85.05.25.4 0.2 epochs
0.5 epochs
1 epochs
2 epochs
5 epochs
10 epochs
20 epochs
(c)Transformer (narrow and shallow) on LM1B
242526272829210211212213214215
Batch Size3.54.04.55.05.5Best Validation Cross Entropy10k steps
20k steps
50k steps
100k steps
200k steps
500k steps
1000k steps
242526272829210211212213214215
Batch Size3.54.04.55.05.5
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
80
|
200k steps
500k steps
1000k steps
242526272829210211212213214215
Batch Size3.54.04.55.05.5
0.2 epochs
0.5 epochs
1 epochs
2 epochs
5 epochs
10 epochs
20 epochs
(d)Transformer (base) on LM1B
Figure 11: Validation error depends on compute budget more than batch size. Plots show
the best validation error subject to budgets of training steps (left column) or training epochs (right
column). Step budgets favor large batch sizes, while epoch budgets favor small batch sizes.
27
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
5. Discussion
Our goals in measuring the e ects of data parallelism on neural network training were
twofold: rst, we hoped to produce actionable advice for practitioners, and second, we
hoped to understand the utility of building systems capable of very high degrees of data
parallelism. Our results indicate that, for idealized data parallel hardware, there is a universal relationship between training time and batch size, but there is dramatic variation in
how well di erent workloads can make use of larger batch sizes. Across all our experiments,
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
81
|
how well di erent workloads can make use of larger batch sizes. Across all our experiments,
increasing the batch size initially reduced the number of training steps needed proportionally. However, depending on the workload, this perfect scaling regime ended anywhere from
a batch size of 24to a batch size of 213. As batch size increases beyond the perfect scaling
regime, there are diminishing returns (where increasing the batch size by a factor of konly
reduces the number of training steps needed by a factor less than k) that end with a maximum useful batch size (where increasing the batch size no longer changes the number of
training steps needed). Once again, the maximum useful batch size is extremely problemdependent and varied between roughly 29and 216in our experiments. Other workloads
may have the region of perfect scaling end at batch sizes even smaller or larger than the
range we observed, as well as having even smaller or larger maximum useful batch sizes.
On the one hand, the possibility that perfect scaling can extend to batch sizes beyond
213for some workloads is good news for practitioners because it suggests that ecient
data-parallel systems can provide extremely large speedups for neural network training.
On the other hand, the wide variation in scaling behavior across workloads is bad news
because any given workload might have a maximum useful batch size well below the limits
of our hardware. Moreover, for a new workload, measuring the training steps needed as
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
82
|
because any given workload might have a maximum useful batch size well below the limits
of our hardware. Moreover, for a new workload, measuring the training steps needed as
a function of batch size and con rming the boundaries of the three basic scaling regimes
requires expensive experiments. In this work, we have only described how to retrospectively
predict the scaling behavior by tuning the optimization metaparameters for every batch
size. Although Golmant et al. (2018) also described the same basic scaling behavior we
found, in their experiments the relationship did not appear consistently across problems,
across error goals, or in out-of-sample error. In light of our own results, the heuristics they
assumed for adjusting the learning rate as a function of batch size are the likely cause of
these inconsistencies, but this explanation only drives home the inconvenience of having to
carefully tune at every new batch size. We were unable to nd reliable support for any
of the previously proposed heuristics for adjusting the learning rate as a function of batch
size. Thus we are forced to recommend that practitioners tune all optimization parameters
anew when they change the batch size or they risk masking the true behavior of the training
procedure.
If the scaling behavior of workloads with respect to batch size has a simple dependence
on properties of the workload, then we might be able to predict the limits of perfect scaling
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
83
|
procedure.
If the scaling behavior of workloads with respect to batch size has a simple dependence
on properties of the workload, then we might be able to predict the limits of perfect scaling
(or the maximum useful batch size) before running extensive experiments. We could then
prioritize workloads to run on specialized hardware or decide whether gaining access to
specialized hardware would be useful for a given workload of interest. On the one hand, our
results are bad news for practitioners because they show that accurate scaling predictions
must depend on a combination of non-obvious properties of the model, optimizer, and data
set. On the other hand, we have a lot of control over the choice of model and optimizer
28
Measuring the Effects of Data Parallelism on Neural Network Training
and there is some indication that they might be responsible for the largest portion of the
variation between workloads. Our results comparing SGD and SGD with momentum (or
Nesterov momentum) show that, at least for the problems we tried, momentum can extend
perfect scaling to much larger batch sizes, o ering clear guidance for practitioners. Other
optimizers, such as KFAC (Martens and Grosse, 2015; Grosse and Martens, 2016; Ba et al.,
2017), or optimization techniques designed speci cally for massively data parallel systems
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
84
|
2017), or optimization techniques designed speci cally for massively data parallel systems
(e.g. Li et al., 2014), might allow perfect scaling to extend much further. Intuitively, it
seems plausible that optimizers that estimate local curvature information might be able to
bene t more from large batches than optimizers that only use gradients.
Although the model seems to have a large e ect on the maximum useful batch size and
the limit of perfect scaling, our results do not give de nitive answers on exactly how to
design models that scale better for a given optimizer and data set. Even when we kept the
model family xed, we observed somewhat inconsistent results from changing the model
width and depth. Chen et al. (2018) suggested that wider models can exploit larger batch
sizes than narrower models, but their theoretical arguments only apply to linear networks
and fully connected networks with a single hidden layer. In contrast, we found that narrower
variants of the Transformer model scaled better to larger batch sizes, although it is unclear
if the same notion of \width" transfers between di erent types of neural networks.
Unlike the model and optimizer, we generally have much less control over the data set.
Unfortunately, properties of the data set also a ect how well training scales in practice.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
85
|
Unlike the model and optimizer, we generally have much less control over the data set.
Unfortunately, properties of the data set also a ect how well training scales in practice.
Our results are equivocal on whether the number of training examples has any e ect, but
changing the data set entirely can certainly change the scaling behavior with respect to
batch size.
Finally, our results at least partially reconcile con
icting stances in the literature on
whether increasing the batch size degrades model quality. Our experiments show that:
1. Any study that only tunes the learning rate for one batch size and then uses a heuristic
to choose the learning rate for other batch sizes (Goyal et al., 2017; Keskar et al., 2017;
Ho er et al., 2017; Lin et al., 2018; Devarakonda et al., 2017; Golmant et al., 2018)
gives a systematic advantage to the batch size used in tuning (as well as nearby batch
sizes). Our results did not show a simple relationship between the optimal learning
rate and batch size that scales inde nitely (see Figures 8 and 21), so the use of simple
heuristics for batch sizes suciently far from the base batch size could very well explain
the degraded solutions and divergent training reported in prior work. Similarly, the
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
86
|
heuristics for batch sizes suciently far from the base batch size could very well explain
the degraded solutions and divergent training reported in prior work. Similarly, the
optimal values of other metaparameters, such as the momentum and learning rate
decay schedule, should not be assumed to remain constant or scale in a simple way as
the batch size increases.
2. Assuming an epoch budget when comparing solution quality between batch sizes (Masters and Luschi, 2018; Goyal et al., 2017; Lin et al., 2018; Devarakonda et al., 2017),
in e ect, limits an investigation to the perfect scaling region of the steps to result vs
batch size curve (see Figure 1). This budget favors smaller batch sizes because they
will perform more optimizer steps for the same number of training examples (see Section 4.8). Certainly, there are situations where an epoch budget is appropriate, but
there may exist budgets just outside the perfect scaling region that can achieve the
same quality solution, and those budgets may still represent a signi cant reduction
29
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
in the number of training steps required. Moreover, even for a xed model and data
set, simply changing the optimizer can signi cantly extend the perfect scaling regime
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
87
|
in the number of training steps required. Moreover, even for a xed model and data
set, simply changing the optimizer can signi cantly extend the perfect scaling regime
to larger batch sizes. For example, Masters and Luschi (2018) found that test performance of ResNet-8 (without batch normalization) on CIFAR-10 with a xed epoch
budget degraded after batch size 16, but considered only plain mini-batch SGD. Our
experiments con rmed that perfect scaling ends at batch size 16 with plain mini-batch
SGD, but using Nesterov momentum extends the perfect scaling regime to batch size
256 (see Figure 1c).
3. Assuming a step budget when comparing solution quality between batch sizes (Ho er
et al., 2017) might favor larger batch sizes because they will see more training examples
for the same number of gradient updates (see Section 4.8). A step budget is likely
sucient for a larger batch size to reach at least the same performance as a smaller
batch size: we never saw the number of steps to reach a goal validation error increase
when the batch size was increased (see Figure 1).
4. Increasing the batch size reduces noise in the gradient estimates (see Equation 4).
However, the noise in updates due to small batches might, in some cases, provide a
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
88
|
4. Increasing the batch size reduces noise in the gradient estimates (see Equation 4).
However, the noise in updates due to small batches might, in some cases, provide a
helpful regularization e ect (Goodfellow et al., 2016; Smith and Le, 2018). Thankfully,
other regularization techniques, such as label smoothing, can replace this e ect (see
Section 4.6). Others have also used regularization techniques, such as data augmentation (Keskar et al., 2017) and L2regularization (Smith and Le, 2018), to eliminate
the \generalization gap" between two batch sizes.
5. Finally, although we do not believe there is an inherent degradation in solution quality
associated with increasing the batch size, depending on the compute budget, it may
become increasingly dicult to nd good values for the metaparameters with larger
batch sizes. Speci cally, increasing the batch size may shrink the region in metaparameter space corresponding to rapid training in terms of epochs (see Figure 9a), as
previously reported by Breuel (2015b). On the other hand, increasing the batch size
may increase the region in metaparameter space corresponding to rapid training in
terms of steps (see Figure 9b).
5.1 Limitations of our experimental protocol
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
89
|
may increase the region in metaparameter space corresponding to rapid training in
terms of steps (see Figure 9b).
5.1 Limitations of our experimental protocol
When interpreting our results, one should keep in mind any limitations of our experimental
protocol. We do not believe any of these limitations are debilitating, and we hope that
describing these potential areas of concern will spur methodological innovation in future
work.
Firstly, we were unable to avoid some amount of human judgment when tuning metaparameters. Although we did not tune metaparameters by hand, we speci ed the search
spaces for automatic tuning by hand and they may not have been equally appropriate for
all batch sizes, despite our best e orts. We are most con dent in our search spaces that
tuned the fewest metaparameters (such as in our experiments that only tuned learning rate
and momentum). We found it quite dicult to be con dent that our tuning was sucient
when we searched over learning rate decay schedules; readers should be aware that the steps
to result measurement is generally quite sensitive to the learning rate schedule. Thus, we
30
Measuring the Effects of Data Parallelism on Neural Network Training
may not have sampled enough trials at some batch sizes or, nearly equivalently, our search
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
90
|
30
Measuring the Effects of Data Parallelism on Neural Network Training
may not have sampled enough trials at some batch sizes or, nearly equivalently, our search
spaces may have been too wide at some batch sizes. Even though we veri ed that the best
trial was not on the boundary of the search space, this by no means guarantees that we
found the globally optimal metaparameters.
Smaller batch sizes typically had more opportunities to measure validation error and,
when validation error was noisy, got more chances to sample a lucky validation error. Batch
sizes (usually larger ones) that did not reach the goal validation error using the rst search
space used revised search spaces that gave them an extra bite of the apple, so to speak.
Finally, our analysis does not consider how robustly we can reach a goal error rate. For
instance, we did not distinguish between batch sizes where all 100 trials achieved the goal
validation error and batch sizes where only one of the 100 trials achieved the goal. The
maximum or minimum value over a set of trials is not usually a very robust statistic, but
something like the 50thpercentile trial mostly reveals information about the search space.
We tried to strike a balance between studying realistic workloads and being able to repeat
our experiments so many times that these uncertainty questions became trivial. Ultimately,
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
91
|
We tried to strike a balance between studying realistic workloads and being able to repeat
our experiments so many times that these uncertainty questions became trivial. Ultimately,
we opted to study realistic workloads and simply report results for the optimal trials.
6. Conclusions and Future Work
Increasing the batch size is a simple way to produce valuable speedups across a range of
workloads, but, for all workloads we tried, the bene ts diminished well within the limits of
current hardware. Unfortunately, blindly increasing the batch size to the hardware limit
will not produce a large speedup for all workloads. However, our results suggest that
some optimization algorithms may be able to consistently extend perfect scaling across
many models and data sets. Future work should perform our same measurements with
other optimizers, beyond the closely-related ones we tried, to see if any existing optimizer
extends perfect scaling across many problems. Alternatively, if we only need speedups for
speci c, high-value problems, we could also consider designing models that extend perfect
scaling to much larger batch sizes. However, unlike the optimizer, practitioners are likely
to tailor their model architectures to the speci c problems at hand. Therefore, instead of
searching for model architectures that happen to scale extremely well, future work should
try to uncover general principles for designing models that can scale perfectly to larger
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
92
|
searching for model architectures that happen to scale extremely well, future work should
try to uncover general principles for designing models that can scale perfectly to larger
batch sizes. Even if such principles remain elusive, we would still bene t from methods
to prospectively predict the scaling behavior of a given workload without requiring careful
metaparameter tuning at several di erent batch sizes. Finally, the deep learning community
can always bene t from methodical experiments designed to test hypotheses, characterize
phenomena, and reduce confusion, to balance more exploratory work designed to generate
new ideas for algorithms and models.
Acknowledgements
We thank Tomer Koren for helpful discussions. We also thank Justin Gilmer and Simon
Kornblith for helpful suggestions and comments on the manuscript. Finally, we thank Matt
J. Johnson for lending us some computing resources.
31
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
Appendix A. Data Set Details
This section contains details of the data sets summarized in Table 1.
A.1 Data Set Descriptions and Pre-Processing
MNIST (LeCun et al., 1998) is a classic handwritten digit image classi cation data set
with 10 mutually exclusive classes. We split the original training set into 55,000 training
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
93
|
with 10 mutually exclusive classes. We split the original training set into 55,000 training
images and 5,000 validation images, and used the ocial test set of 10,000 images. We did
not use data augmentation.
Fashion MNIST (Xiao et al., 2017) is another reasonably simple image classi cation
data set with 10 mutually exclusive classes. It was designed as a drop-in replacement for
MNIST. We split the original training set into 55,000 training images and 5,000 validation
images, and used the ocial test set of 10,000 images. We did not use data augmentation.
CIFAR-10 (Krizhevsky, 2009) is an image classi cation data set of 32 32 color images
with 10 mutually exclusive classes. We split the original training set into 45,000 training
images and 5,000 validation images. We used the ocial test set of 10,000 images. We
pre-processed each image by subtracting the average value across all pixels and channels
and dividing by the standard deviation.15We did not use data augmentation.
ImageNet (Russakovsky et al., 2015) is an image classi cation data set with 1,000
mutually exclusive classes. We split the ocial training set into 1,281,167 training images
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
94
|
mutually exclusive classes. We split the ocial training set into 1,281,167 training images
and 50,045 test images, and used the ocial validation set of 50,000 images. We preprocessed the images and performed data augmentation in a similar way to Simonyan and
Zisserman (2014). Speci cally, at training time, we sampled a random integer S2[256;512],
performed an aspect-preserving resize so that the smallest side had length S, and took a
random crop of size (224 ;224). We randomly re
ected the images horizonally, but unlike
Simonyan and Zisserman (2014), we did not distort the colors. At evaluation time, we
performed an aspect-preserving resize so that the smallest side had length 256, and took a
central crop of size (224 ;224). In both training and evaluation, we then subtracted the global
mean RGB value from each pixel using the values computed by Simonyan and Zisserman
(2014).16
Open Images v4 (Krasin et al., 2017) is a data set of 9 million images that are
annotated with image-level labels and object bounding boxes.17The image labels were
generated by a computer vision model and then veri ed as either positive ornegative labels
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
95
|
annotated with image-level labels and object bounding boxes.17The image labels were
generated by a computer vision model and then veri ed as either positive ornegative labels
by human annotators. We only considered the 7,186 \trainable" classes with at least 100
human-annotated positives in the training set. We ltered the ocial subsets by keeping
only images with at least one positive trainable label, which produced training, validation
and test sets of size 4,526,492; 41,225; and 124,293 images, respectively. On average, each
image in the training set has 2.9 human-annotated positive labels, while each image in the
validation and test sets have 8.4 human-annotated positive labels. We only considered the
human-annotated positives and assumed all other classes were negative. We pre-processed
the images and performed data augmentation identically to ImageNet.
15. We used the TensorFlow op tf.image.per image standardization .
16. See https://gist.github.com/ksimonyan/211839e770f7b538e2d8#description for the mean RGB values used.
17. Available at https://storage.googleapis.com/openimages/web/index.html .
32
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
96
|
17. Available at https://storage.googleapis.com/openimages/web/index.html .
32
Measuring the Effects of Data Parallelism on Neural Network Training
LM1B (Chelba et al., 2014) is a text data set of English news articles.18We used
the ocial training set and created validation and test sets using les news.en.heldout00000-of-00050 and news.en.heldout-00001-of-00050 , respectively. These splits contain 30,301,028; 6,075; and 6,206 sentences, respectively. We used an invertable word
tokenizer to split the text into sub-word tokens with a vocabulary of size 32,000.19On
average, the training set contains around 20 tokens per sentence and the validation and test
sets contain around 29 tokens per sentence. At training time, we clipped long sentences
to the rst 64 tokens, which a ected only about 2% of sentences. We did not clip long
sentences at evaluation time. The maximum sentence across the validation and test sets
has 476 tokens.
Common Crawl is a repository of web data containing over 3 billion web pages.20
We ltered and processed the data set identically to Anil et al. (2018).21The vocabulary
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
97
|
We ltered and processed the data set identically to Anil et al. (2018).21The vocabulary
contains 24,006 sub-word tokens. We randomly partitioned the sentences into a training
set (99.98%) and a holdout set (0.02%). Our training set contains 25:8 billion sentences.
We used the rst 6,075 sentences of the holdout set as our validation set, which is the same
number of sentences in our LM1B validation set. Some sentences are tens of thousands of
tokens long. To maintain consistency with our LM1B processing, we clipped sentences to
64 tokens at training time and 476 at evaluation time.
A.2 Evaluation Metrics
We use classi cation error for MNIST, Fashion MNIST, CIFAR-10, and ImageNet. To
compute this metric, we consider the model's classi cation for each image to be the class it
assigns the highest probability. Then
classi cation error =# incorrect classi cations
# classi cations:
We use class-agnostic average precision (AP) for Open Images. To compute this
metric, we rst rank each image-class pair by the predicted likelihood of the class being a
true positive for that image. Then
AP=1
wnmX
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
98
|
metric, we rst rank each image-class pair by the predicted likelihood of the class being a
true positive for that image. Then
AP=1
wnmX
k=1Precision(k)Relevance(k); (7)
where Precision( k) is the precision when considering the top kimage-class pairs, Relevance( k)
is an indicator function equal to 1 if the kthimage-class pair is a veri ed positive and 0
otherwise,nis the number of images in the validation set, mis the number of classes, and
wis the number of positive labels. Average precision was proposed for Open Images by Veit
et al. (2017). Due to false negatives in the validation set, Veit et al. (2017) only computed
APover the the human-annotated classes in each image. However, on average, each image
18. Available at http://www.statmt.org/lm-benchmark/ .
19. The code for processing the raw data and generating the vocabulary is available at https://github.
com/tensorflow/tensor2tensor/blob/master/tensor2tensor/data_generators/lm1b.py
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
99
|
com/tensorflow/tensor2tensor/blob/master/tensor2tensor/data_generators/lm1b.py
20. Available at http://commoncrawl.org/2017/07/june-2017-crawl-archive-now-available/ .
21. See https://github.com/google-research/google-research/tree/master/codistillation for document IDs.
33
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
in the validation set only has 8.4 positive and 4 negative human-annotated classes, so each
image is only evaluated over 12 classes out of 7,186. This yields misleadingly high values
ofAP. Instead, we compute APover all classes in each image, which may underestimate
the true APdue to false negatives in the validation set, but is more indicative of the true
performance in our experience. We compute APusing an ecient approximation of the
area under the discrete precision-recall curve.22
We use average per-token cross entropy error for LM1B and Common Crawl. For a
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
100
|
area under the discrete precision-recall curve.22
We use average per-token cross entropy error for LM1B and Common Crawl. For a
single sentence s= (w1;:::;wm), letp(wjjw1;:::;wj 1) denote the model's predicted probability of the token wjgiven all prior tokens in the sentence. Thus, the predicted logprobability of sis logp(s) =Pm
j=1logp(wjjw1;:::;wj 1). We compute the average per-token
cross entropy error over a data set fs1;:::;sngas
cross entropy error =Pn
i=1logp(sn)Pn
i=1len(sn);
where len(s) denotes the number of tokens in s. This is the logarithm of the per-token
perplexity.
Appendix B. Model Details
In this section we give the architectural details of the models summarized in Table 2.
In addition to the descriptions below, each model has a task-speci c output layer. Models
trained on MNIST, Fashion MNIST, CIFAR-10, and ImageNet (classi cation with mutually
exclusive labels) use a softmax output layer to model the probability distribution over
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
101
|
exclusive labels) use a softmax output layer to model the probability distribution over
classes. Models trained on Open Images (classi cation with multiple labels per image) use
a sigmoid output layer to model the probability of each class. Models trained on LM1B and
Common Crawl (language modeling) use a softmax output layer to model the probability
of the next word in a sentence given all prior words in the sentence.
Fully Connected is a fully connected neural network with ReLU activation function.
Hidden layers use dropout with probability 0.4 during training. We vary the number of
layers and number of units per layer in di erent experiments to investigate the impact of
model size. We use the notation FC- N1-...-Nkto denote a fully connected neural network
withkhidden layers and Niunits in the ithlayer.
Simple CNN consists of 2 convolutional layers with max-pooling followed by 1 fully
connected hidden layer. The convolutional layers use 5 5 lters with stride length 1,
\same" padding (Goodfellow et al., 2016), and ReLU activation function. Max pooling uses
22 windows with stride length 2. The fully connected layer uses dropout with probability
0.4 during training. We used three di erent model sizes: base has 32 and 64 lters in the
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
102
|
0.4 during training. We used three di erent model sizes: base has 32 and 64 lters in the
convolutional layers and 1,024 units in the fully connected layer; narrow has 16 and 32
lters in the convolutional layers and 512 units in the fully connected layer; and wide has
64 and 128 lters in the convolutional layers and 2,048 units in the fully connected layer.
We used the base model unless otherwise speci ed.
22. Equation 7 can be interpreted as a right Riemann sum of the discrete precision-recall curve f(ri;pi)ji=
1;:::;wg, whereri=i=w andpiis the maximum precision among all values of precision with recallri(each value of recall may correspond to di erent values of precision at di erent classi cation
thresholds). We use the TensorFlow op tf.metrics.auc with curve="PR" ,numthresholds=200 , and
summation method="careful interpolation" .
34
Measuring the Effects of Data Parallelism on Neural Network Training
ResNet-8 consists of 7 convolutional layers with residual connections followed by 1 fully
connected hidden layer. We used the model described in section 4.2 of He et al. (2016a) with
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
103
|
connected hidden layer. We used the model described in section 4.2 of He et al. (2016a) with
n= 1, but with the improved residual block described by He et al. (2016b). We removed
batch normalization, which is consistent with Masters and Luschi (2018).
ResNet-50 consists of 49 convolutional layers with residual connections followed by
1 fully connected hidden layer. We used the model described in section 4.1 of He et al.
(2016a), but with the improved residual block described by (He et al., 2016b). We replaced
batch normalization (Io e and Szegedy, 2015) with ghost batch normalization to keep the
training objective xed between batch sizes and to avoid possible negative e ects from
computing batch normalization statistics over a large number of examples (Ho er et al.,
2017). We used a ghost batch size of 32 for all experiments. We also applied label smoothing
(Szegedy et al., 2016) to regularize the model at training time, which was helpful for larger
batch sizes. The label smoothing coecient was a metaparameter that we tuned in our
experiments.
VGG-11 consists of 8 convolutional layers followed by 3 fully connected hidden layers.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
104
|
experiments.
VGG-11 consists of 8 convolutional layers followed by 3 fully connected hidden layers.
We used the model referred to as \model A" by Simonyan and Zisserman (2014).
LSTM is a one hidden-layer LSTM model (Hochreiter and Schmidhuber, 1997). It is
a simpler variant of the LSTM-2048-512 model described by Jozefowicz et al. (2016), with
1,024 embedding dimensions, 2,048 hidden units, and 512 projection dimensions. We did
not use bias parameters in the output layer because we found this improved performance
in our preliminary experiments.
Transformer is a self-attention model that was originally presented for machine translation (Vaswani et al., 2017). We used it as an autoregressive language model by applying
the decoder directly to the sequence of word embeddings for each sentence. We used four
di erent sizes: the base model described by Vaswani et al. (2017); a shallow model that
is identical to the base model except with only two hidden layers instead of six; a narrow
and shallow model that is identical to the shallow model except with half as many hidden
units and attention heads as well as half the lter size; and a wide model that is identical
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
105
|
and shallow model that is identical to the shallow model except with half as many hidden
units and attention heads as well as half the lter size; and a wide model that is identical
to the base model except with double the number of hidden units and attention heads as
well as double the lter size. We used the base model unless otherwise speci ed.
Appendix C. Learning Rate Schedules
We chose our learning rate schedule by experimenting with a variety of di erent schedules
for ResNet-50 on ImageNet. For each schedule, we speci ed the following metaparameters:
0: initial learning rate
: decay factor ( >0)
T: number of training steps until the learning rate decays from 0to 0
Each schedule corresponds to a decay function d(t), such that the learning rate at
training step tis
(t) =(
d(t)0iftT;
0 ift>T:
We experimented with the following decay functions:
35
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
Constant :d(t) = 1
Linear :d(t) = 1 (1 )t
T
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
106
|
Constant :d(t) = 1
Linear :d(t) = 1 (1 )t
T
Cosine (Loshchilov and Hutter, 2017): d(t) = +(1 )
2
1 + cost
T
Exponential Polynomial :d(t) = + (1 )
1 t
T, where>0
Inverse Exponential Polynomial :d(t) =
+(1 )(t
T), where>0
Exponential :d(t) = t=T
We also tried piecewise linear learning rate schedules. These schedules are speci ed by
a sequence of pairs f(t0;0);:::;(tk;k)g, with 0 = t0< t1::: < tk, such that the learning
rate at training step tis
(t) =(
i+i+1 i
ti+1 ti(t ti) iftit<ti+1;
k ifttk:
The schedules used by both He et al. (2016a) (piecewise constant) and Goyal et al. (2017)
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
107
|
k ifttk:
The schedules used by both He et al. (2016a) (piecewise constant) and Goyal et al. (2017)
(linear warm-up followed by piecewise constant) for ResNet-50 on ImageNet can both be
expressed as piecewise linear.
We ran experiments with ResNet-50 on ImageNet, using Nesterov momentum with batch
size 1,024 for 150,000 training steps, while tuning the momentum and all metaparameters
governing the learning rate schedule. We used quasi-random metaparameter search as
described in Section 4. For piecewise linear schedules, we tried 1, 3, and 5 decay events. We
found that it was possible to get good results with several of the schedules we tried, and it is
likely that other schedules would also work well. Ultimately, we chose linear decay because
it performed at least well as all other schedules we tried, while also being the simplest and
requiring only two additional metaparameters.
36
Measuring the Effects of Data Parallelism on Neural Network Training
Appendix D. Additional Plots
2022242628210212214216
Batch Size2022242628210212214216StepsValidation, 0.01
Train, 0.01
(a) Simple CNN on MNIST
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
108
|
Batch Size2022242628210212214216StepsValidation, 0.01
Train, 0.01
(a) Simple CNN on MNIST
242526272829210211212213214215
Batch Size2829210211212213214215216217218219220StepsValidation, 3.9
Train, 3.9 (b) Transformer on LM1B
26272829210211212213214215216
Batch Size210211212213214215216217218219220221StepsValidation, 0.25
Train, 0.15 (c) ResNet-50 on ImageNet
Figure 12: Steps to result on the training set is almost the same as on the validation set.
The evaluation metrics are described in Appendix A.2. Error goals are speci ed in the plot legends.
10-410-310-210-1
Learning Rate (η)0.40.50.60.70.80.91.01.11.21.3Steps to Result1e6 Batch Size 16
10-410-310-210-1
Learning Rate (η)0.51.01.52.02.53.0Steps to Result1e5 Batch Size 128
10-410-310-210-1
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
109
|
Learning Rate (η)0.51.01.52.02.53.0Steps to Result1e5 Batch Size 128
10-410-310-210-1
Learning Rate (η)012345678Steps to Result1e4 Batch Size 2048
10-410-310-210-1
Learning Rate (η)0.81.01.21.41.61.82.02.22.42.6Steps to Result1e4 Batch Size 16384
10-410-310-210-1100
1 - Momentum (1−γ)0.40.50.60.70.80.91.01.11.21.3Steps to Result1e6 Batch Size 16
10-410-310-210-1100
1 - Momentum (1−γ)0.51.01.52.02.53.0Steps to Result1e5 Batch Size 128
10-410-310-210-1100
1 - Momentum (1−γ)012345678Steps to Result1e4 Batch Size 2048
10-410-310-210-1100
1 - Momentum (1−γ)0.81.01.21.41.61.82.02.22.42.6Steps to Result1e4 Batch Size 16384
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
110
|
Figure 13: Validating metaparameter search spaces for Transformer on LM1B. Rows correspond to the metaparameters we tuned (learning rate and momentum
) and columns correspond
to di erent batch sizes. The x-axis is the search range that was sampled by the quasi-random search
algorithm. Blue dots represent trials that reached the goal of 3.9 validation cross entropy error, and
yellow stars correspond to trials that achieved the goal in the fewest steps. We deem these search
spaces appropriate because the yellow stars are not on the boundaries.
37
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
10-410-310-210-1100101
Learning Rate (η)1.21.41.61.82.02.22.4Steps to Result1e6 Batch Size 64
10-410-310-210-1100101
Learning Rate (η)0.80.91.01.11.21.31.41.5Steps to Result1e5 Batch Size 1024
10-410-310-210-1100101
Learning Rate (η)1.21.31.41.51.61.71.81.92.0Steps to Result1e4 Batch Size 8192
10-210-1100101
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
111
|
10-210-1100101
Learning Rate (η)0.550.600.650.700.750.800.850.900.951.00Steps to Result1e4 Batch Size 65536
10-410-310-210-1
1 - Momentum (1−γ)1.21.41.61.82.02.22.4Steps to Result1e6 Batch Size 64
10-410-310-210-1
1 - Momentum (1−γ)0.80.91.01.11.21.31.41.5Steps to Result1e5 Batch Size 1024
10-410-310-210-1
1 - Momentum (1−γ)1.21.31.41.51.61.71.81.92.0Steps to Result1e4 Batch Size 8192
10-410-310-210-1
1 - Momentum (1−γ)0.550.600.650.700.750.800.850.900.951.00Steps to Result1e4 Batch Size 65536
10-410-310-210-1
Decay Factor (α)1.21.41.61.82.02.22.4Steps to Result1e6 Batch Size 64
10-410-310-210-1
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
112
|
10-410-310-210-1
Decay Factor (α)0.80.91.01.11.21.31.41.5Steps to Result1e5 Batch Size 1024
10-410-310-210-1
Decay Factor (α)1.21.31.41.51.61.71.81.92.0Steps to Result1e4 Batch Size 8192
10-410-310-210-1
Decay Factor (α)0.550.600.650.700.750.800.850.900.951.00Steps to Result1e4 Batch Size 65536
1.2 1.4 1.6 1.8 2.0 2.2 2.4
Decay Steps (T) 1e61.21.41.61.82.02.22.4Steps to Result1e6 Batch Size 64
0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5
Decay Steps (T) 1e50.80.91.01.11.21.31.41.5Steps to Result1e5 Batch Size 1024
1.0 1.2 1.4 1.6 1.8
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
113
|
1.0 1.2 1.4 1.6 1.8
Decay Steps (T) 1e41.01.21.41.61.8Steps to Result1e4 Batch Size 8192
0.5 0.6 0.7 0.8 0.9 1.0
Decay Steps (T) 1e40.50.60.70.80.91.0Steps to Result1e4 Batch Size 65536
0.02
0.02 0.06 0.10 0.14
Label Smoothing1.21.41.61.82.02.22.4Steps to Result1e6 Batch Size 64
0.02
0.02 0.06 0.10 0.14
Label Smoothing0.80.91.01.11.21.31.41.5Steps to Result1e5 Batch Size 1024
0.02
0.02 0.06 0.10 0.14
Label Smoothing1.21.31.41.51.61.71.81.92.0Steps to Result1e4 Batch Size 8192
0.02
0.02 0.06 0.10 0.14
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
114
|
0.02
0.02 0.06 0.10 0.14
Label Smoothing0.550.600.650.700.750.800.850.900.951.00Steps to Result1e4 Batch Size 65536
Figure 14: Validating metaparameter search spaces for ResNet-50 on ImageNet. Rows
correspond to the metaparameters we tuned (initial learning rate 0, momentum
, learning rate
decay parameters ,T, and label smoothing parameter) and columns correspond to di erent batch
sizes. For all parameters except the label smoothing parameter, the x-axis is the search range sampled by the quasi-random search algorithm. The label smoothing parameter was sampled uniformly
inf0;0:01;0:1gforb214andf0;0:1gforb>214. Blue dots represent trials that reached the goal
validation error rate of 0.25, and yellow stars correspond to trials that achieved the goal in the fewest
steps. We deem these search spaces appropriate because the yellow stars are not on the boundaries.
38
Measuring the Effects of Data Parallelism on Neural Network Training
2022242628210212214216
Batch Size2526272829210211212213214215216217218StepsSteps to Reach 0.03 Validation Error
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
115
|
2022242628210212214216
Batch Size2526272829210211212213214215216217218StepsSteps to Reach 0.03 Validation Error
FC-1024-1024-1024
Simple CNN
(a) Fully Connected vs Simple CNN on MNIST
2526272829210211212213214215216
Batch Size210211212213214215216217218219220221StepsSteps to Reach 0.35 Validation Error
ResNet-50
VGG-11 (b) ResNet-50 vs VGG-11 on ImageNet
242526272829210211212213214215
Batch Size210211212213214215216217218219220221StepsSteps to Reach 3.9 Validation Cross Entropy
Transformer
LSTM
(c) Transformer vs LSTM on LM1B
2022242628210212214216
Batch Size2526272829210211212213214215216217218StepsSteps to Reach 0.03 Validation Error
FC-1024
FC-128-128-128
FC-256-256-256
FC-512-512-512
FC-1024-1024-1024
FC-2048-2048-2048 (d) Fully Connected sizes on MNIST
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
116
|
FC-512-512-512
FC-1024-1024-1024
FC-2048-2048-2048 (d) Fully Connected sizes on MNIST
2022242628210212214216
Batch Size23242526272829210211212213214215216StepsSteps to Reach 0.01 Validation Error
Simple CNN
Simple CNN Narrow
Simple CNN Wide
(e) Simple CNN sizes on MNIST
242526272829210211212213214215
Batch Size210211212213214215216217218219220221StepsSteps to Reach 4.2 Validation Cross Entropy
Wide
Base
Shallow
Narrow and Shallow (f) Transformer sizes on LM1B
Figure 15: Figure 3 without the y-axis normalized.
39
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
2022242628210212214216
Batch Size2022242628210212214216StepsMNIST, 0.01
Fashion MNIST, 0.1
(a) Simple CNN on di erent data sets
26272829210211212213214215216
Batch Size210211212213214215216217218219220221StepsImageNet, 0.25
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
117
|
26272829210211212213214215216
Batch Size210211212213214215216217218219220221StepsImageNet, 0.25
Open Images, 0.31 (b) ResNet-50 on di erent data sets
242526272829210211212213214215
Batch Size2829210211212213214215216217218219220Steps
Common Crawl, 3.9
LM1B, 3.9 (c) Transformer on di erent data sets
Figure 16: Figure 5 without the y-axis normalized.
2022242628210212214216
Batch Size242526272829210211212213214Steps100% of Images, 0.02
50% of Images, 0.02
25% of Images, 0.02
12.5% of Images, 0.02
(a) Simple CNN on MNIST subsets
26272829210211212213214215216
Batch Size210211212213214215216217218219220221Steps100% of Images, 0.25
50% of Images, 0.3
50% of Classes, 0.3 (b) ResNet-50 on ImageNet subsets
Figure 17: Figure 6 without the y-axis normalized.
0 2000 4000 6000 8000 10000
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
118
|
50% of Classes, 0.3 (b) ResNet-50 on ImageNet subsets
Figure 17: Figure 6 without the y-axis normalized.
0 2000 4000 6000 8000 10000
Training Steps0.00.20.40.60.81.0Training Error
Label Smoothing = 0.00
Label Smoothing = 0.10
0 2000 4000 6000 8000 10000
Training Steps0.20.30.40.50.60.70.80.91.0Validation Error
Label Smoothing = 0.00
Label Smoothing = 0.10
Figure 18: Label smoothing reduces over tting at large batch sizes. Plots are training
curves for the two best models with and without label smoothing for ResNet-50 on ImageNet with
batch size 216. The two models correspond to di erent metaparameter tuning trials, so the learning
rate, Nesterov momentum, and learning rate schedule were independently chosen for each trial. The
two trials shown are those that reached the highest validation error at any point during training, for
label smoothing equal to 0 and 0 :1 respectively.
40
Measuring the Effects of Data Parallelism on Neural Network Training
2022242628210212214216
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
119
|
label smoothing equal to 0 and 0 :1 respectively.
40
Measuring the Effects of Data Parallelism on Neural Network Training
2022242628210212214216
Batch Size0.00300.00350.00400.00450.00500.0055Validation ErrorBest Validation Error Per Batch Size
Label Smoothing = 0.00
Label Smoothing = 0.10
(a) Simple CNN on MNIST
2022242628210212214216
Batch Size0.0560.0580.0600.0620.0640.0660.0680.070Validation ErrorBest Validation Error Per Batch Size
Label Smoothing = 0.00
Label Smoothing = 0.10 (b) Simple CNN on Fashion MNIST
Figure 19: Label smoothing helps all batch sizes for Simple CNN on MNIST and Fashion
MNIST. There is no consistent trend of label smoothing helping smaller or larger batch sizes more.
Each point corresponds to a di erent metaparameter tuning trial, so the learning rate, Nesterov
momentum, and learning rate schedule are independently chosen for each point. The training budget
is xed for each batch size, but varies between batch sizes.
10-510-410-310-210-1100101
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
120
|
is xed for each batch size, but varies between batch sizes.
10-510-410-310-210-1100101
Learning rate / (1 -Momentum)3.03.54.04.55.05.56.06.57.07.5Validation Cross EntropyBatch Size 16
10-510-410-310-210-1100101
Learning rate / (1 -Momentum)Batch Size 128
10-510-410-310-210-1100101
Learning rate / (1 -Momentum)Batch Size 2048
10-510-410-310-210-1100101
Learning rate / (1 -Momentum)Batch Size 16384
(a)Transformer on LM1B
10-310-210-1100101102103104
Learning rate / (1 - Momentum)0.00.20.40.60.81.01.2Validation ErrorBatch Size 64
10-310-210-1100101102103104
Learning rate / (1 - Momentum)Batch Size 1024
10-310-210-1100101102103104
Learning rate / (1 - Momentum)Batch Size 8192
10-310-210-1100101102103104
Learning rate / (1 - Momentum)Batch Size 65536
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
121
|
10-310-210-1100101102103104
Learning rate / (1 - Momentum)Batch Size 65536
(b)ResNet-50 on ImageNet
Figure 20: Validation error vs e ective learning rate. Training budgets are consistent for each
batch size, but not between batch sizes. These plots are projections of the entire metaparameter
search space, which is 2-dimensional for Transformer on LM1B (see Figure 13) and 5-dimensional
for ResNet-50 on ImageNet (see Figure 14).
41
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
2022242628210212214216
Batch Size2-112-92-72-52-32-121232527Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic
(a) Simple CNN on MNIST
2123252729211213215217
Batch Size2-102-82-62-42-220222426Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (b) Simple CNN on Fashion MNIST
212223242526272829210211212213
Batch Size2-132-112-92-72-52-32-12123Learning RateOptimal Learning Rate
Linear Heuristic
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
122
|
Batch Size2-132-112-92-72-52-32-12123Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (c) ResNet-8 on CIFAR-10
26272829210211212213214215216
Batch Size2-92-82-72-62-52-42-32-22-120212223Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic
(d) ResNet-50 on ImageNet
26272829210211212213214215
Batch Size2-112-102-92-82-72-62-52-42-32-22-120Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (e) ResNet-50 on Open Images
242526272829210211212213214215
Batch Size2-102-92-82-72-62-52-42-32-22-120212223Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (f) Transformer on LM1B
2526272829210211212213214
Batch Size2-102-92-82-72-62-52-42-32-22-12021Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
123
|
Linear Heuristic
Square Root Heuristic
(g) Transformer on Common Crawl
2526272829210211212213214215216
Batch Size2-142-132-122-112-102-92-82-72-62-52-42-32-2Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (h) VGG-11 on ImageNet
242526272829210211212213214215
Batch Size2-152-142-132-122-112-102-92-82-72-62-52-42-3Learning RateOptimal Learning Rate
Linear Heuristic
Square Root Heuristic (i) LSTM on LM1B
Figure 21: Optimal learning rates do not always follow linear or square root scaling
heuristics. Learning rates correspond to the trial that reached the goal validation error in the fewest
training steps (see Figure 1). For models using learning rate decay schedules (ResNet-8, ResNet-50,
VGG-11), plots are based on the initial learning rate. See Figure 22 for the corresponding plot of
optimal momentum, and Figure 8 for the corresponding plot of e ective learning rate.
42
Measuring the Effects of Data Parallelism on Neural Network Training
2022242628210212214216
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
124
|
42
Measuring the Effects of Data Parallelism on Neural Network Training
2022242628210212214216
Batch Size0.650.700.750.800.850.900.951.00Momentum
(a) Simple CNN on MNIST
2123252729211213215217
Batch Size0.920.930.940.950.960.970.980.991.00Momentum (b) Simple CNN on Fashion MNIST
212223242526272829210211212213
Batch Size0.600.650.700.750.800.850.900.951.00Momentum (c) ResNet-8 on CIFAR-10
26272829210211212213214215216
Batch Size0.940.950.960.970.980.991.00Momentum
(d) ResNet-50 on ImageNet
26272829210211212213214215
Batch Size0.900.920.940.960.981.00Momentum (e) ResNet-50 on Open Images
242526272829210211212213214215
Batch Size0.880.900.920.940.960.981.00Momentum (f) Transformer on LM1B
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
125
|
Batch Size0.880.900.920.940.960.981.00Momentum (f) Transformer on LM1B
2526272829210211212213214
Batch Size0.930.940.950.960.970.980.991.00Momentum
(g) Transformer on Common Crawl
2526272829210211212213214215216
Batch Size0.9650.9700.9750.9800.9850.9900.9951.000Momentum (h) VGG-11 on ImageNet
242526272829210211212213214215
Batch Size0.940.960.981.001.021.04Momentum (i) LSTM on LM1B*
Figure 22: Optimal momentum has no consistent relationship with batch size. Momentum
corresponds to the trial that reached the goal validation error in the fewest training steps (see
Figure 1). See Figure 21 for the corresponding plot of optimal learning rate, and Figure 8 for the
corresponding plot of e ective learning rate. *For LSTM on LM1B, we only tuned with xed
= 0:99.
43
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
126
|
corresponding plot of e ective learning rate. *For LSTM on LM1B, we only tuned with xed
= 0:99.
43
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
Step budget Epoch budget
2123252729211213215
Batch Size0.0030.0040.0050.0060.0070.008Best Validation Error2k steps
5k steps
10k steps
50k steps
100k steps
500k steps
1000k steps
2000k steps
5000k steps
2123252729211213215
Batch Size0.0030.0040.0050.0060.0070.0085 epochs
10 epochs
20 epochs
50 epochs
100 epochs
200 epochs
500 epochs
1000 epochs
10000 epochs
(a)Simple CNN on MNIST: Validation Error
2123252729211213215
Batch Size0.0040.0050.0060.0070.0080.0090.010Best Test Error2k steps
5k steps
10k steps
50k steps
100k steps
500k steps
1000k steps
2000k steps
5000k steps
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
127
|
5k steps
10k steps
50k steps
100k steps
500k steps
1000k steps
2000k steps
5000k steps
2123252729211213215
Batch Size0.0040.0050.0060.0070.0080.0090.010
5 epochs
10 epochs
20 epochs
50 epochs
100 epochs
200 epochs
500 epochs
1000 epochs
10000 epochs
(b)Simple CNN on MNIST: Test Error
Figure 23: Zoomed version of Figure 11a.
Step budget Epoch budget
2123252729211213215
Batch Size0.0500.0550.0600.0650.0700.0750.080Best Validation Error5k steps
10k steps
20k steps
50k steps
100k steps
200k steps
500k steps
5000k steps
2123252729211213215
Batch Size0.0500.0550.0600.0650.0700.0750.080
10 epochs
20 epochs
50 epochs
100 epochs
200 epochs
500 epochs
1000 epochs
10000 epochs
(a)Simple CNN on Fashion MNIST: Validation Error
2123252729211213215
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
128
|
200 epochs
500 epochs
1000 epochs
10000 epochs
(a)Simple CNN on Fashion MNIST: Validation Error
2123252729211213215
Batch Size0.0600.0650.0700.0750.080Best Test Error5k steps
10k steps
20k steps
50k steps
100k steps
200k steps
500k steps
5000k steps
2123252729211213215
Batch Size0.0600.0650.0700.0750.080
10 epochs
20 epochs
50 epochs
100 epochs
200 epochs
500 epochs
1000 epochs
10000 epochs
(b)Simple CNN on Fashion MNIST: Test Error
Figure 24: Zoomed version of Figure 11b.
44
Measuring the Effects of Data Parallelism on Neural Network Training
References
Martin Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Je rey Dean,
Matthieu Devin, Sanjay Ghemawat, Geo rey Irving, Michael Isard, et al. TensorFlow: a
system for large-scale machine learning. In Conference on Operating Systems Design and
Implementation , volume 16, pages 265{283. USENIX, 2016.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
129
|
system for large-scale machine learning. In Conference on Operating Systems Design and
Implementation , volume 16, pages 265{283. USENIX, 2016.
Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch SGD: Training ResNet-50 on ImageNet in 15 minutes. arXiv preprint arXiv:1711.04325 , 2017.
Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E. Dahl, and
Geo rey E. Hinton. Large scale distributed neural network training through online
distillation. In International Conference on Learning Representations , 2018. URL
https://openreview.net/forum?id=rkr1UDeC- .
Jimmy Ba, Roger Grosse, and James Martens. Distributed second-order optimization using
Kronecker-factored approximations. In International Conference on Learning Representations , 2017. URL https://openreview.net/forum?id=SkkTMpjex .
L eon Bottou and Olivier Bousquet. The tradeo s of large scale learning. In Advances in
Neural Information Processing Systems , pages 161{168, 2008.
Olivier Bousquet, Sylvain Gelly, Karol Kurach, Olivier Teytaud, and Damien Vincent.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
130
|
Neural Information Processing Systems , pages 161{168, 2008.
Olivier Bousquet, Sylvain Gelly, Karol Kurach, Olivier Teytaud, and Damien Vincent.
Critical hyper-parameters: No random, no cry. arXiv preprint arXiv:1706.03200 , 2017.
Thomas M Breuel. Benchmarking of LSTM networks. arXiv preprint arXiv:1508.02774 ,
2015a.
Thomas M Breuel. The e ects of hyperparameters on SGD training of neural networks.
arXiv preprint arXiv:1508.02788 , 2015b.
Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn,
and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling. In Conference of the International Speech Communication Association ,
2014.
Jianmin Chen, Xinghao Pan, Rajat Monga, Samy Bengio, and Rafal Jozefowicz. Revisiting distributed synchronous SGD. In International Conference on Learning
Representations Workshop Track , 2016. URL https://openreview.net/forum?id=
D1VDZ5kMAu5jEJ1zfEWL .
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
131
|
Representations Workshop Track , 2016. URL https://openreview.net/forum?id=
D1VDZ5kMAu5jEJ1zfEWL .
Lingjiao Chen, Hongyi Wang, Jinman Zhao, Dimitris Papailiopoulos, and Paraschos
Koutris. The e ect of network width on the performance of large-batch training. arXiv
preprint arXiv:1806.03791 , 2018.
Valeriu Codreanu, Damian Podareanu, and Vikram Saletore. Scale out for large minibatch
SGD: Residual network training on ImageNet-1K with improved accuracy and reduced
time to train. arXiv preprint arXiv:1711.04291 , 2017.
45
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
Aditya Devarakonda, Maxim Naumov, and Michael Garland. AdaBatch: Adaptive batch
sizes for training deep neural networks. arXiv preprint arXiv:1712.02029 , 2017.
Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can
generalize for deep nets. In International Conference on Machine Learning , pages 1019{
1028, 2017.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
132
|
generalize for deep nets. In International Conference on Machine Learning , pages 1019{
1028, 2017.
Noah Golmant, Nikita Vemuri, Zhewei Yao, Vladimir Feinberg, Amir Gholami, Kai
Rothauge, Michael W Mahoney, and Joseph Gonzalez. On the computational ineciency
of large batch sizes for stochastic gradient descent. arXiv preprint arXiv:1811.12941 ,
2018.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning . MIT Press, 2016.
URL http://www.deeplearningbook.org .
Priya Goyal, Piotr Doll ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD:
training ImageNet in 1 hour. arXiv preprint arXiv:1706.02677 , 2017.
Roger Grosse and James Martens. A Kronecker-factored approximate Fisher matrix for
convolution layers. In International Conference on Machine Learning , pages 573{582,
2016.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
133
|
Roger Grosse and James Martens. A Kronecker-factored approximate Fisher matrix for
convolution layers. In International Conference on Machine Learning , pages 573{582,
2016.
Elad Hazan. Introduction to online convex optimization. Foundations and Trends in Optimization , 2(3-4):157{325, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In Conference on Computer Vision and Pattern Recognition , pages 770{778.
IEEE, 2016a.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep
residual networks. In European Conference on Computer Vision , pages 630{645. Springer,
2016b.
Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan
Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling
is predictable, empirically. arXiv preprint arXiv:1712.00409 , 2017.
Geo rey Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine
learning, lecture 6a: overview of mini-batch gradient descent, 2012. URL https:
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
134
|
learning, lecture 6a: overview of mini-batch gradient descent, 2012. URL https:
//www.cs.toronto.edu/ ~tijmen/csc321/slides/lecture_slides_lec6.pdf .
Sepp Hochreiter and J urgen Schmidhuber. Long short-term memory. Neural Computation ,
9(8):1735{1780, 1997.
Elad Ho er, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the
generalization gap in large batch training of neural networks. In Advances in Neural
Information Processing Systems , pages 1731{1741, 2017.
46
Measuring the Effects of Data Parallelism on Neural Network Training
Sergey Io e and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning ,
pages 448{456, 2015.
Prateek Jain, Sham M. Kakade, Rahul Kidambi, Praneeth Netrapalli, and Aaron Sidford.
Parallelizing stochastic gradient descent for least squares regression: Mini-batching, averaging, and model misspeci cation. Journal of Machine Learning Research , 18(223):1{42,
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
135
|
2018. URL http://jmlr.org/papers/v18/16-595.html .
Norman P Jouppi, Cli Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder
Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In International Symposium on Computer
Architecture , pages 1{12. IEEE, 2017.
Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring
the limits of language modeling. arXiv preprint arXiv:1602.02410 , 2016.
Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of Fisher information in deep neural networks: Mean eld approach. arXiv preprint arXiv:1806.01316 ,
2018.
Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping
Tak Peter Tang. On large-batch training for deep learning: Generalization gap and
sharp minima. In International Conference on Learning Representations , 2017. URL
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
136
|
Tak Peter Tang. On large-batch training for deep learning: Generalization gap and
sharp minima. In International Conference on Learning Representations , 2017. URL
https://openreview.net/forum?id=H1oyRlYgg .
Rahul Kidambi, Praneeth Netrapalli, Prateek Jain, and Sham M. Kakade. On the insuciency of existing momentum schemes for stochastic optimization. In International
Conference on Learning Representations , 2018. URL https://openreview.net/forum?
id=rJTutzbA- .
Jack Kiefer, Jacob Wolfowitz, et al. Stochastic estimation of the maximum of a regression
function. The Annals of Mathematical Statistics , 23(3):462{466, 1952.
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In
International Conference on Learning Representations , 2015.
Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina
Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci,
Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
137
|
Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen
Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy.
OpenImages: A public dataset for large-scale multi-label and multi-class image classi cation., 2017. URL https://storage.googleapis.com/openimages/web/index.html .
Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. URL http://www.cs.toronto.edu/ ~kriz/
learning-features-2009-TR.pdf .
Guanghui Lan. An optimal method for stochastic composite optimization. Mathematical
Programming , 133(1-2):365{397, 2012.
47
Shallue, Lee, Antognini, Sohl-Dickstein, Frostig, and Dahl
Yann Le Cun, L eon Bottou, Genevieve B. Orr, and Klaus-Robert M uller. Ecient backprop.
InNeural Networks, Tricks of the Trade , Lecture Notes in Computer Science LNCS 1524.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
138
|
InNeural Networks, Tricks of the Trade , Lecture Notes in Computer Science LNCS 1524.
Springer Verlag, 1998. URL http://leon.bottou.org/papers/lecun-98x .
Yann LeCun, Corinna Cortes, and CJ Burges. MNIST handwritten digit database, 1998.
URL http://yann.lecun.com/exdb/mnist .
Yann LeCun, Yoshua Bengio, and Geo rey Hinton. Deep learning. Nature , 521(7553):436,
2015.
Mu Li, Tong Zhang, Yuqiang Chen, and Alexander J Smola. Ecient mini-batch training
for stochastic optimization. In International Conference on Knowledge Discovery and
Data Mining , pages 661{670. ACM, 2014.
Tao Lin, Sebastian U Stich, and Martin Jaggi. Don't use large mini-batches, use local SGD.
arXiv preprint arXiv:1808.07217 , 2018.
Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm
restarts. In International Conference on Learning Representations , 2017. URL https:
//openreview.net/forum?id=Skq89Scxx .
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
139
|
restarts. In International Conference on Learning Representations , 2017. URL https:
//openreview.net/forum?id=Skq89Scxx .
Siyuan Ma, Raef Bassily, and Mikhail Belkin. The power of interpolation: Understanding
the e ectiveness of SGD in modern over-parametrized learning. In International Conference on Machine Learning , pages 3331{3340, 2018.
James Martens and Roger Grosse. Optimizing neural networks with Kronecker-factored
approximate curvature. In International Conference on Machine Learning , pages 2408{
2417, 2015.
Dominic Masters and Carlo Luschi. Revisiting small batch training for deep neural networks.
arXiv preprint arXiv:1804.07612 , 2018.
Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of
convergence O(1=k2). In Doklady AN USSR , volume 269, pages 543{547, 1983.
Boris T Polyak. Some methods of speeding up the convergence of iteration methods. USSR
Computational Mathematics and Mathematical Physics , 4(5):1{17, 1964.
Herbert Robbins and Sutton Monro. A stochastic approximation method. The Annals of
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
140
|
Computational Mathematics and Mathematical Physics , 4(5):1{17, 1964.
Herbert Robbins and Sutton Monro. A stochastic approximation method. The Annals of
Mathematical Statistics , 22(3):400{407, 1951.
David E Rumelhart, Geo rey E Hinton, and Ronald J Williams. Learning representations
by back-propagating errors. Nature , 323(6088):533, 1986.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large
scale visual recognition challenge. International Journal of Computer Vision , 115(3):
211{252, 2015.
48
Measuring the Effects of Data Parallelism on Neural Network Training
Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From foundations to algorithms . Cambridge University Press, 2014. URL https://books.google.
com/books?id=OE9etAEACAAJ .
Ohad Shamir. Without-replacement sampling for stochastic gradient methods. In Advances
in Neural Information Processing Systems , pages 46{54, 2016.
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
141
|
Ohad Shamir. Without-replacement sampling for stochastic gradient methods. In Advances
in Neural Information Processing Systems , pages 46{54, 2016.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556 , 2014.
Samuel L. Smith and Quoc V. Le. A Bayesian perspective on generalization and stochastic
gradient descent. In International Conference on Learning Representations , 2018. URL
https://openreview.net/forum?id=BJij4yg0Z .
Ilya Sutskever, James Martens, George E. Dahl, and Geo rey E. Hinton. On the importance
of initialization and momentum in deep learning. In International Conference on Machine
Learning , pages 1139{1147, 2013.
Christian Szegedy, Vincent Vanhoucke, Sergey Io e, Jon Shlens, and Zbigniew Wojna.
Rethinking the Inception architecture for computer vision. In Conference on Computer
Vision and Pattern Recognition , pages 2818{2826. IEEE, 2016.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
142
|
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances
in Neural Information Processing Systems , pages 5998{6008, 2017.
Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge J Belongie.
Learning from noisy large-scale datasets with minimal supervision. In Conference on
Computer Vision and Pattern Recognition , pages 6575{6583. IEEE, 2017.
D Randall Wilson and Tony R Martinez. The general ineciency of batch training for
gradient descent learning. Neural Networks , 16(10):1429{1451, 2003.
Yuhuai Wu, Mengye Ren, Renjie Liao, and Roger Grosse. Understanding short-horizon
bias in stochastic meta-optimization. In International Conference on Learning Representations , 2018. URL https://openreview.net/forum?id=H1MczcgR- .
Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a novel image dataset for
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1811.03600
|
143
|
Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a novel image dataset for
benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 , 2017.
Dong Yin, Ashwin Pananjady, Max Lam, Dimitris Papailiopoulos, Kannan Ramchandran,
and Peter Bartlett. Gradient diversity: a key ingredient for scalable distributed learning.
InInternational Conference on Arti cial Intelligence and Statistics , 2018. URL http:
//proceedings.mlr.press/v84/yin18a.html .
Yang You, Zhao Zhang, Cho-Jui Hsieh, James Demmel, and Kurt Keutzer. ImageNet
training in minutes. arXiv preprint arXiv:1709.05011 , 2017.
49
|
1811.03600
|
Measuring the Effects of Data Parallelism on Neural Network Training
|
Recent hardware developments have dramatically increased the scale of data
parallelism available for neural network training. Among the simplest ways to
harness next-generation hardware is to increase the batch size in standard
mini-batch neural network training algorithms. In this work, we aim to
experimentally characterize the effects of increasing the batch size on
training time, as measured by the number of steps necessary to reach a goal
out-of-sample error. We study how this relationship varies with the training
algorithm, model, and data set, and find extremely large variation between
workloads. Along the way, we show that disagreements in the literature on how
batch size affects model quality can largely be explained by differences in
metaparameter tuning and compute budgets at different batch sizes. We find no
evidence that larger batch sizes degrade out-of-sample performance. Finally, we
discuss the implications of our results on efforts to train neural networks
much faster in the future. Our experimental data is publicly available as a
database of 71,638,836 loss measurements taken over the course of training for
168,160 individual models across 35 workloads.
|
http://arxiv.org/pdf/1811.03600
|
[
"Christopher J. Shallue",
"Jaehoon Lee",
"Joseph Antognini",
"Jascha Sohl-Dickstein",
"Roy Frostig",
"George E. Dahl"
] |
[
"cs.LG",
"stat.ML"
] | null |
Journal of Machine Learning Research 20 (2019) 1-49
|
cs.LG
|
20181108
|
20190719
|
[
{
"id": "1806.03791"
},
{
"id": "1712.02029"
},
{
"id": "1811.12941"
},
{
"id": "1804.07612"
},
{
"id": "1808.07217"
},
{
"id": "1602.02410"
},
{
"id": "1708.07747"
},
{
"id": "1712.00409"
},
{
"id": "1711.04291"
},
{
"id": "1709.05011"
},
{
"id": "1508.02774"
},
{
"id": "1711.04325"
},
{
"id": "1706.03200"
},
{
"id": "1806.01316"
},
{
"id": "1508.02788"
},
{
"id": "1811.03600"
},
{
"id": "1706.02677"
}
] |
1801.04871
|
0
|
Building a Conversational Agent Overnight with Dialogue Self-Play
Pararth Shah1, Dilek Hakkani-T ¨ur, Gokhan T ¨ur, Abhinav Rastogi,
Ankur Bapna, Neha Nayak, Larry Heck
Google AI
Mountain View, CA, USA
Abstract
We propose Machines Talking To Machines
(M2M), a framework combining automation
and crowdsourcing to rapidly bootstrap endto-end dialogue agents for goal-oriented dialogues in arbitrary domains. M2M scales to
new tasks with just a task schema and an API
client from the dialogue system developer, but
it is also customizable to cater to task-specific
interactions. Compared to the Wizard-of-Oz
approach for data collection, M2M achieves
greater diversity and coverage of salient dialogue flows while maintaining the naturalness
of individual utterances. In the first phase, a
simulated user bot and a domain-agnostic system bot converse to exhaustively generate dialogue “outlines”, i.e. sequences of template
utterances and their semantic parses. In the
second phase, crowd workers provide contextual rewrites of the dialogues to make the utterances more natural while preserving their
meaning. The entire process can finish within
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
1
|
second phase, crowd workers provide contextual rewrites of the dialogues to make the utterances more natural while preserving their
meaning. The entire process can finish within
a few hours. We propose a new corpus of
3,000 dialogues spanning 2 domains collected
with M2M, and present comparisons with popular dialogue datasets on the quality and diversity of the surface forms and dialogue flows.
1 Introduction
Goal-oriented dialogue agents trained using supervised learning methods work best when trained on
dialogues of the same task. However, when developing a dialogue agent to assist a user for completing a new task, for example scheduling a doctors appointment via an online portal, a dataset of
human-agent dialogues for that task may not be
available since no dialogue agent exists for interacting with that particular API. One popular approach is to collect and annotate free-form dialogues via crowdsourcing using a Wizard-of-Oz
1Correspondence to [email protected]
Figure 1: Our proposed M2M framework: (1) the dialogue developer provides a task schema and an API
client, (2) automated bots generate dialogue outlines,
(3) crowd workers rewrite the utterances and validate
slot spans, (4) a dialogue model is trained with supervised learning on the dataset. The whole process can
complete in under 8 hours.
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
2
|
slot spans, (4) a dialogue model is trained with supervised learning on the dataset. The whole process can
complete in under 8 hours.
setup (Wen et al. (2016); Asri et al. (2017)). However, this is an expensive and lossy process as the
free-form dialogues collected from crowdworkers
(i) might not cover all the interactions that the
agent is expected to handle, (ii) might contain dialogues unfit for use as training data (for instance if
the crowd workers use language that is either too
simplistic or too convoluted), and (iii) may havearXiv:1801.04871v1 [cs.AI] 15 Jan 2018
errors in dialogue act annotations, requiring an expensive manual filtering and cleaning step by the
dialogue developer.
Another approach, popular among consumerfacing voice assistants, is to enable third-party
developers to build dialogue “experiences” or
“skills” focusing on individual tasks (e.g. DialogFlow1, Alexa Skills2, wit.ai3). This provides
the dialogue developer with full control over how
a particular task is handled, allowing her to incrementally add new features to that experience.
However, this approach relies heavily on the developer to engineer every aspect of the conversational interaction and anticipate all ways in which
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
3
|
However, this approach relies heavily on the developer to engineer every aspect of the conversational interaction and anticipate all ways in which
users might interact with the agent for completing
that task. It is desirable to expand this approach to
make it more data-driven, bringing it closer to the
Wizard-of-Oz approach popular in the dialogue research community.
We present Machines Talking To Machines
(M2M), a functionality-driven process for training
dialogue agents. The primary goal is to reduce the
cost and effort required to build dialogue datasets
by automating the task-independent steps so that
a dialogue developer is required to provide only
the task-specific aspects of the dialogues. Another
goal is to obtain a higher quality of dialogues in
terms of: (i) diversity of language as well as dialogue flows, (ii) coverage of all expected user behaviors, and (iii) correctness of supervision labels.
Finally, this framework is aimed towards bootstrapping dialogue agents up to the point where
they can be deployed to serve real users with an
acceptable task completion rate, after which they
should be improved directly from user feedback
using reinforcement learning.
Previous work for building semantic parsers
(Wang et al. (2015)) and parsers for mapping
natural language questions to structured queries
(Zhong et al. (2017)) rely on crowd sourcing
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
4
|
(Wang et al. (2015)) and parsers for mapping
natural language questions to structured queries
(Zhong et al. (2017)) rely on crowd sourcing
to map automatically generated structured representations to single-shot natural language utterances. However, generating multi-turn dialogues
in this manner requires co-ordination among multiple participating agents. Inspired by recent
AI game-playing literature (Silver et al. (2016,
2017)), we introduce a notion of “dialogue selfplay” where two or more conversational agents in1https://dialogflow.com
2https://developer.amazon.com/alexa-skills-kit
3https://wit.aiteract by choosing discrete conversational actions
to exhaustively generate dialogue histories. In this
work, we employ an agenda-based user simulator
agent (Schatzmann et al. (2007)) and a finite state
machine based system agent for the self-play step.
In Section 2 we describe the mechanics of M2M
and in Sections 3 and 4 we discuss the user simulation and crowdsourcing aspects of our method. In
Section 5 we present datasets collected with this
framework that we are releasing with this paper
and in Section 6 we evaluate our approach by comparing our datasets with popular dialogue datasets.
We conclude with a discussion of related work in
Section 7.
2 M2M
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
5
|
and in Section 6 we evaluate our approach by comparing our datasets with popular dialogue datasets.
We conclude with a discussion of related work in
Section 7.
2 M2M
At a high level (Figure 1), M2M connects a developer , who provides the task-specific information, and a framework , which provides the taskindependent information, for generating dialogues
centered around completing the task. Formally,
the framework Fmaps a task specification Tto
a set of dialogues D:
F(T)!D=fdi; i21: : : Ng (1)
di= [(ui
1; : : : ; ui
ni);(ai
1; : : : ; ai
ni)](2)
Each dialogue diis a sequence of natural language utterances (or dialogue turns )ui
jand their
corresponding annotations ai
j. A dialogue turn annotation ai
jincludes the semantic parse of that turn
as well as additional information tied to that turn,
for example who spoke at that turn and the dialogue state at that point in the dialogue.
2.1 Dialogue task specification
The input to the framework is a task specification obtained from the dialogue developer, which
defines the scope of the dialogue interactions for
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
6
|
The input to the framework is a task specification obtained from the dialogue developer, which
defines the scope of the dialogue interactions for
the task in question. Dialogue agents can be employed to complete a wide variety of tasks. In this
work we focus on database querying applications,
which involve a relational database which contains
entities that the user would like to browse and select through a natural language dialogue. This formulation covers a large variety of tasks that are
expected of automated dialogue agents, including
all tasks that map to filling a form and executing
a transaction on some website. The attributes of
the entities (i.e. columns of the database) induce
Figure 2: Example of generating an outline and its paraphrase. See text for details.
a schema Sof “slots”. Each slot could be a constraint that the user cares about when selecting an
entity. The developer must provide an API client
Cwhich can be queried with a SQL-like syntax to
return a list of matching candidate entities for any
valid combination of slot values. The task schema
and API client together form the task specification,
T= (S; C). (Figure 2a.)
Dialogues that involve procedural turn-taking
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
7
|
and API client together form the task specification,
T= (S; C). (Figure 2a.)
Dialogues that involve procedural turn-taking
(for example reading out a recipe or playing textbased games), or deal with unstructured knowledge (for example question answering over a text
document), are among tasks that are not covered
by this formulation. These classes of dialogues
can be handled by modifying the self-play phase
of the framework to generate outlines for these dialogue types.
2.2 Outline generation via self-play
With the task specification, the framework must
generate a set of dialogues centered around that
task. We divide this into two separate steps, F=
F2F1, where F1maps the task specification to
a set of outlines O, andF2maps each outline to a
natural language dialogue:
F1(T)!O=foi; i21: : : Ng (3)
oi= [(ti
1; : : : ; ti
ni);(ai
1; : : : ; ai
ni)] (4)
F2(foig)!fdig (5)
We define an outline oias a sequence of template utterances ti
jand their corresponding annotationsai
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
8
|
F2(foig)!fdig (5)
We define an outline oias a sequence of template utterances ti
jand their corresponding annotationsai
j. Template utterances are simplistic statements with language that is easy to generate with
a few rules, as we will describe below. An outline encapsulates the flow of the dialogue while
abstracting out the variation in natural language ofthe surface forms of each dialogue turn. Outlines
are easier to generate using self-play between a
user bot and a system bot as the bots do not need to
generate complex and diverse language that mimics real users and assistants.
To generate an outline, the framework first samples a scenario from the task specification. We
define a scenario as a user profile and user goals,
si= (pi; gi)(Figure 2b). In a goal-oriented dialogue, the user wants to accomplish goals with
the assistance of the dialogue agent, for example
booking movie tickets or reserving restaurant tables. Each goal is associated with constraints that
map to slots of the schema, for example the movie
name, genre, number of tickets, and price range.
The slots in a user goal can have fixed values (e.g.
genre should be “comedy” or the user will deny
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
9
|
The slots in a user goal can have fixed values (e.g.
genre should be “comedy” or the user will deny
the offer), a list of possible values (e.g. genre
should “comedy” or “action”), flexible values (e.g.
“comedy” is preferred, but the user is open to other
options), or open values (e.g. the user is open to
seeing movies of any genre). In the multi-domain
setting, a goal’s slot values can refer to previous
goals, for example the user may want to buy a
movie ticket and then get dinner after the movie at
a restaurant near the theatre chosen in the preceding sub-dialogue. A scenario generator samples
goals gifrom the task specification by randomly
choosing the constraint type and values for every
slot in the schema. The values are chosen from a
set that includes all available values in the database
as well as some non-existent values to create unsatisfiable user goals.
In addition to the user goal, the flow of the dialogue is also dependent on the personality of the
user. A user could be verbose in specifying more
constraints in a single turn, or could prefer to give
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
10
|
user. A user could be verbose in specifying more
constraints in a single turn, or could prefer to give
each constraint separately. Similarly a user could
be more or less amenable to changing their goal
if their original constraints are not satisfiable. We
define a user profile vector pito encapsulate all
the task-independent characteristics of the user’s
behavior. In its simplest version, picould be modeled as a vector of probabilities concerning independent aspects of the user’s behavior, which
could be passed into a programmed user simulator. Alternatively, picould be an embedding of a
user profile in a latent space, which could condition a learned user simulator model. In our setup,
the scenario generator samples pifrom a manually
specified distribution of typical user profiles.
With the dialogue scenario si, the framework
performs dialogue self-play between a user bot
BUandsystem bot BSto generate a sequence of
turn annotations ai
1: : : ai
nias follows:
BU=P(ai
jjai
1; : : : ; ai
j 1; pi; gi) (6)
BS=P(ai
j+1jai
1; : : : ; ai
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
11
|
j 1; pi; gi) (6)
BS=P(ai
j+1jai
1; : : : ; ai
j; S; C ) (7)
Each turn annotation ai
jconsists of a dialogue
frame that encodes the semantics of the turn as
adialogue act and a slot-value map , for example “inform(date=tomorrow, time=evening)” is a
dialogue frame that informs the system of the
user’s constraints for the date and time slots. Table 4 in Appendix A has a full list of dialogue
acts.BUmaps a (possibly empty) dialogue history
ai
1: : : ai
j 1and a scenario pi; gito a distribution
over turn annotations for the next user turn. Similarly, BSmaps a dialogue history, task schema S
and API client Cto a distribution over system turn
annotations. In dialogue self-play (Figure 2c), a
new turn annotation ai
jis iteratively sampled from
each bot until either the user’s goals are achieved
and the user exits the dialogue with a “bye()” act,
or a maximum number of turns are reached.
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
12
|
each bot until either the user’s goals are achieved
and the user exits the dialogue with a “bye()” act,
or a maximum number of turns are reached.
In our setup, BUis an agenda-based user simulator (Schatzmann et al. (2007)) with a modification that the action selection model is conditioned on the user profile in addition to the user
goal and dialogue history. BSis modeled as a finite state machine (Hopcroft et al. (2006)) which
encodes a set of task-independent rules for constructing system turns, with each turn consisting
of a response frame which responds to the user’s
previous turn, and an initiate frame which drives
the dialogue forward through a predetermined sequence of sub-dialogues. For database querying
applications, these sub-dialogues are: gather userpreferences, query a database via an API, offer
matching entities to the user, allow user to modify
preferences or request more information about an
entity, and finally complete the transaction (buying or reserving the entity). By exploring a range
of parameter values for BUandBSand sampling
a large number of outlines, dialogue self-play can
generate a diverse set of dialogue outlines for the
task.
Finally, a template utterance generator maps
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
13
|
a large number of outlines, dialogue self-play can
generate a diverse set of dialogue outlines for the
task.
Finally, a template utterance generator maps
each turn annotation to a template utterance, ai
j!
ti
j, using a domain-general grammar similar to the
one described in Wang et al. (2015). Alternatively,
the developer can provide a list of templates to use
for some or all of the dialogue frames, for example
if they want more control over the language used
in the system turns. The template utterances ti
jare
an important bridge between the turn annotation
ai
jand the corresponding natural language utteranceui
j, since crowd workers may not understand
the annotations if presented in symbolic form.
2.3 Crowdsourced paraphrases
To obtain a natural language dialogue from its outline,F2(oi)!di, the framework employs crowd
sourcing to paraphrase template utterances ti
jinto
more natural sounding utterances ui
j. The paraphrase task is designed as a “contextual rewrite”
task where a crowd worker sees the full dialogue
template ti
1: : : ti
ni, and provides the natural language utterances ui
1: : : ui
nifor each turn of the dialogue. A screenshot of the contextual rewrite task
interface is provided in Figure 3. This encourages
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
14
|
1: : : ui
nifor each turn of the dialogue. A screenshot of the contextual rewrite task
interface is provided in Figure 3. This encourages
the crowd worker to inject linguistic phenomena
like coreference (“Reserve that restaurant”) and
lexical entrainment into the utterances. We show
the same outline to K > 1crowdworkers to get
more diversity of natural language utterances for
the same dialogue, fui
jgk.
Since ti
jandui
jare paraphrases of each other,
the annotations ai
jautomatically apply to ui
j, eliminating the need for an expensive annotation step.
In practice, for a fraction of the utterances, the
automatic annotation does not succeed either due
to crowd workers not following instructions properly or if the utterance contains a paraphrase of
a slot value, for example when the crowd worker
rephrases “between 5pm and 8pm” as “some time
in the evening”. We employ a second round of
crowdsourcing for validating the utterances. For
eachui
j, we ask two crowd workers if it has the
same meaning as the corresponding template ti
j,
and we drop the utterance if either of the crowd
workers say no. Dialogues which end up having
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
15
|
same meaning as the corresponding template ti
j,
and we drop the utterance if either of the crowd
workers say no. Dialogues which end up having
no natural language utterance for at least one of
the turns are dropped from the dataset. For the
remaining utterances, slot values from the annotationai
jare tagged in the utterance with substring
match. If a slot value cannot be found automatically, we show it to two crowd workers and ask
them to annotate the slot span. Alternatively, such
annotation errors be detected and corrected by active learning (Hakkani-T ¨ur et al. (2002); Tur et al.
(2003)).
2.4 Dataset expansion
The rewrites t!fugkcollected via the crowdsourcing step F2can be compiled into a map
L(a)! fugk. As an optional step, this map
could be leveraged to synthetically expand the
dataset beyond what is economically feasible to
collect via crowdsourcing. The self-play step F1
can be executed to generate a larger set of outlines OS>> O . For each turn annotation ai
j
ofoi2OS, a natural language utterance is sampled4fromL(ai
j)to create the corresponding dialogue di2DS>> D . Dialogues in the synthetic
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
16
|
j
ofoi2OS, a natural language utterance is sampled4fromL(ai
j)to create the corresponding dialogue di2DS>> D . Dialogues in the synthetic
setDScould have utterances that were written by
crowdworkers under a different context, so these
dialogues are of a slightly lower quality.
2.5 Model training
The dialogues di2D(orDS) have natural language turns along with annotations of dialogue
acts, slot spans, dialogue state and API state for
each turn. These labels are sufficient for training dialogue models from recent literature: either
component-wise models for language understanding (Bapna et al. (2017)), state tracking (Rastogi
et al. (2017)), dialogue policy (Shah et al. (2016))
and language generation (Nayak et al. (2017)), or
end-to-end models (Wen et al. (2016)). Further,
we can construct a natural language user simulator by combining UBwithL(a), and use it to
train end-to-end dialogue models with reinforcement learning (Liu et al. (2017)).
3 User simulation and dialogue self-play
Our framework hinges on having a generative
model of a user that is reasonably close to actual
users of the system. The motivation is that while
4Ifai
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
17
|
Our framework hinges on having a generative
model of a user that is reasonably close to actual
users of the system. The motivation is that while
4Ifai
j=2L, then oiis dropped from OS.it is hard to develop precise models of user behavior customized for every type of dialogue interaction, it is possible to create a domain-general
user simulator that operates at a higher level of
abstraction (dialogue acts) and encapsulates common patterns of user behavior for a broad class of
dialogue tasks. Seeding the user simulator with a
task-specific schema of intents, slot names and slot
values allows the framework to generate a variety
of dialogue flows tailored to that specific task. Developing a general user simulator targeting a broad
class of tasks, for example database querying applications, has significant leverage as adding a new
conversational pattern to the simulator benefits the
outlines generated for dialogue interfaces to any
database or third-party API.
Another concern with the use of a user simulator is that it restricts the generated dialogue
flows to only those that are engineered into the
user model. In comparison, asking crowd workers to converse without any restrictions could generate interesting dialogues that are not anticipated
by the dialogue developer. Covering complex interactions is important when developing datasets
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
18
|
user model. In comparison, asking crowd workers to converse without any restrictions could generate interesting dialogues that are not anticipated
by the dialogue developer. Covering complex interactions is important when developing datasets
to benchmark research aimed towards building
human-level dialogue systems. However, we argue that for consumer-facing chatbots, the primary
aim is reliable coverage of critical user interactions. Existing methods for developing chatbots
with engineered finite state machines implicitly
define a model of expected user behavior in the
states and transitions of the system agent. A user
simulator makes this user model explicit and is a
more systematic approach for a dialogue developer to reason about the user behaviors handled
by the agent. Similarly, having more control over
the dialogue flows present in the dataset ensures
that all and only expected user and system agent
behaviors are present in the dataset. Our crowd
sourcing setup obtains diverse natural language realizations of the abstract dialogue flows generated
via self-play. A dialogue agent bootstrapped with
such a dataset can be deployed in front of users
with a guaranteed minimum task completion rate.
Subsequently, the dialogue agent can be directly
improved from real user interactions, for which
crowdsourcing is anyways an imperfect substitute.
The self-play step also uses a system bot BS
that generates valid system turns for a given task.
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
19
|
crowdsourcing is anyways an imperfect substitute.
The self-play step also uses a system bot BS
that generates valid system turns for a given task.
Since our framework uses a rule-based bot which
takes user dialogue acts as inputs and emits a neural network based dialogue agent that works with
natural language utterances, the framework effectively distills expert knowledge into a learned neural network. The developer can customize the behavior of the neural agent by modifying the component rules of BS. Further, the cost of developing
BScan be amortized over a large class of dialogue
tasks by building a domain-agnostic bot for handling a broad task like database querying applications, similar to US. Finally, in contrast to a rulebased bot, a neural dialogue agent is amenable to
further improvement from direct user interactions
via reinforcement learning (Su et al. (2016); Liu
et al. (2017)), opening up the possibility of lifelong improvement in the quality of the dialogue
agent.
4 Crowdsourcing
In the Wizard-of-Oz setting, the dialogue task
specification is used to construct tasks by sampling slot values from the API client. A task is
then shown to a pair of crowd workers who are
asked to converse in natural language to complete
the task. Subsequently, the collected dialogues
are manually annotated with dialogue act and slot
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
20
|
asked to converse in natural language to complete
the task. Subsequently, the collected dialogues
are manually annotated with dialogue act and slot
span labels for training dialogue models. This
process is expensive as the two annotation tasks
given to crowd workers in the WOz setting are
difficult and therefore time consuming: identifying the dialogue acts of an utterance requires understanding the precise meaning of each dialogue
act, and identifying all slot spans in an utterance
requires checking the utterance against all slots in
the schema. As a result, the crowdsourced annotations may need to be cleaned by an expert. In
contrast, M2M significantly reduces the crowdsourcing expense by automatically annotating a
majority of the dialogue turns and annotating the
remaining turns with two simpler crowdsourcing
tasks, “Does this utterance contain this particular
slot value?” and “Do these two utterances have the
same meaning?”, which are more efficiently done
by an average crowd worker.
Further, the lack of control over crowd workers’ behavior in the Wizard-of-Oz setting can lead
to dialogues that may not reflect the behavior of
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
21
|
Further, the lack of control over crowd workers’ behavior in the Wizard-of-Oz setting can lead
to dialogues that may not reflect the behavior of
real users, for example if the crowd worker provides all constraints in a single turn. Such lowquality dialogues either need to be manually removed from the dataset or the crowd participants
need to be given additional instructions or train-Dataset Slots Train Dev Test
Restaurantprice range, location,
restaurant name, category,
num people, date, time1116 349 775
Movietheatre name, movie, date,
time, num people384 120 264
Table 1: Dialogues collected with M2M.
ing to encourage better interactions (Asri et al.
(2017)). M2M avoids this issue by using dialogue
self-play to systematically generate all usable dialogue outlines, and simplifying the crowdsourcing
step to a dialogue paraphrase task.
5 Datasets
We are releasing5two datasets totaling 3000 dialogues collected using M2M for the tasks of buying a movie ticket and reserving a restaurant table. (Table 1). The datasets were collected by
first generating outlines using dialogue self-play
and then rewriting the template utterances using
crowd sourcing.
6 Evaluations
We present some experiments with the M2M
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
22
|
first generating outlines using dialogue self-play
and then rewriting the template utterances using
crowd sourcing.
6 Evaluations
We present some experiments with the M2M
datasets to evaluate the M2M approach for collecting dialogue datasets and training conversational
agents with that data.
6.1 Dialogue diversity
First we will investigate the claim that M2M leads
to higher coverage of dialogue features in the
dataset. We compare the M2M Restaurants training dialogues with the DSTC2 (Henderson et al.
(2013)) training set which also deals with restaurant reservations (Table 2). M2M compares favorably to DSTC2 on the ratio of unique unigrams and bigrams to total number of tokens in
the dataset, which signifies a greater variety of
surface forms as opposed to repeating the same
words and phrases. Similarly, we count the number of unique “transitions” at the semantic frame
level, defined as a pair of annotations ai; ai+1of
contiguous turns. This gives a measure of diversity of dialogue flows in the dataset. M2M has
3x the number of unique transitions per turn of
the dataset. We also count unique “subdialogues”,
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
23
|
3x the number of unique transitions per turn of
the dataset. We also count unique “subdialogues”,
i.e. sequences of transitions ai; ai+1; : : : ; a i+kfor
k=f3;5g, and observe that M2M has fewer repetitions of subdialogues compared to DSTC2.
5https://github.com/google-research-datasets/simulateddialogue
MetricDSTC2
(Train)M2M Rest.
(Train)
Dialogues 1611 1116
Total turns 11670 6188
Total tokens 199295 99932
Avg. turns per dialogue 14.49 11.09
Avg. tokens per turn 8.54 8.07
Unique tokens /
Total tokens0.0049 0.0092
Unique bigrams /
Total tokens0.0177 0.0670
Unique transitions /
Total turns0.0982 0.2646
Unique subdialogues(k=3) /
Total subdialogues(n=3)0.1831 0.3145
Unique subdialogues(k=5) /
Total subdialogues(n=5)0.5621 0.7061
Unique full outlines /
Total dialogues0.9243 0.9292
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
24
|
Total subdialogues(n=5)0.5621 0.7061
Unique full outlines /
Total dialogues0.9243 0.9292
Table 2: Comparing DSTC2 and M2M Restaurants
datasets on diversity of language and dialogue flows.
6.2 Human evaluation of dataset quality
For a subjective evaluation of the quality of the
M2M datasets, we ran an experiment showing the
final dialogues to crowd workers and asking them
to rate each user and system turn between 1 to 5
on multiple dimensions. Figure 4 in the Appendix
presents the interface shown to crowd workers for
collecting the ratings. Each turn was shown to 3
crowd workers. Table 3 presents the mean and
standard deviation of ratings aggregated over all
turns of the datasets.
7 Related work and discussion
We presented M2M, an extensible framework
for rapidly bootstrapping goal-oriented conversational agents. Comparisons with the popular Dialog State Tracking Challenge 2 dataset (Henderson
et al. (2013)) show that M2M can be leveraged for
rapidly creating high-quality datasets for training
conversational agents in arbitrary domains. A key
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
25
|
et al. (2013)) show that M2M can be leveraged for
rapidly creating high-quality datasets for training
conversational agents in arbitrary domains. A key
benefit of our framework is that it is fully controllable via multiple knobs: the task schema, the scenario generator, the user profile and behavior, the
system policy and the template generator. PyDial
(Ultes et al. (2017)), an extensible open-source
toolkit which provides domain-independent implementations of dialogue system modules, could
be extended to support M2M by adding dialogue
self-play functionality.
The user and system bots in this work are imple-M2M
RestaurantsM2M
Movies
User:
Natural 4.66 (0.54) 4.70 (0.49)
System:
Polite 4.23 (0.62) 4.27 (0.62)
Clear 4.72 (0.52) 4.75 (0.48)
Optimal 4.26 (0.76) 4.32 (0.75)
Table 3: Human evaluation of dialogues collected with
M2M. Average of crowd worker scores (from 1 to 5) for
user and system turns (standard deviation in brackets).
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
26
|
Table 3: Human evaluation of dialogues collected with
M2M. Average of crowd worker scores (from 1 to 5) for
user and system turns (standard deviation in brackets).
mented using task-general rules so that any transactional or form-filling task could be handled with
only the task schema. For more complex tasks,
the developer can extend the user and system bots
or the canonical utterance generator by adding
their own rules. These components could also
be replaced by machine learned generative models if available. Task Completion Platform (TCP)
(Crook et al. (2016)) introduced a task configuration language for building goal-oriented dialogue
interactions. The state update and policy modules
of TCP could be used to implement bots that generate outlines for complex tasks.
ParlAI (Miller et al. (2017)), a dialogue research
software platform, provides easy integration with
crowd sourcing for data collection and evaluation.
However, the crowd sourcing tasks are open-ended
and may result in lower quality dialogues as described in Section 4. The crowd sourcing tasks in
M2M are configured to be at a suitable difficulty
level for crowd workers as they are neither openended nor too restrictive. The crowd workers are
asked to paraphrase utterances instead of coming
up with completely new ones.
Acknowledgements
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
27
|
level for crowd workers as they are neither openended nor too restrictive. The crowd workers are
asked to paraphrase utterances instead of coming
up with completely new ones.
Acknowledgements
We thank Georgi Nikolov, Amir Fayazi, Anna
Khasin and Grady Simon for valuable support in
design, implementation and evaluation of M2M.
References
Layla El Asri, Hannes Schulz, Shikhar Sharma,
Jeremie Zumer, Justin Harris, Emery Fine, Rahul
Mehrotra, and Kaheer Suleman. 2017. Frames: A
corpus for adding memory to goal-oriented dialogue
systems. arXiv preprint arXiv:1704.00057 .
Ankur Bapna, Gokhan Tur, Dilek Hakkani-Tur, and
Larry Heck. 2017. Sequential dialogue context
modeling for spoken language understanding. In
Proc. of SIGDIAL .
PA Crook, A Marin, V Agarwal, K Aggarwal, T Anastasakos, R Bikkula, D Boies, A Celikyilmaz,
S Chandramohan, Z Feizollahi, et al. 2016. Task
completion platform: A self-serve multi-domain
goal oriented dialogue platform. NAACL HLT 2016
page 47.
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
28
|
completion platform: A self-serve multi-domain
goal oriented dialogue platform. NAACL HLT 2016
page 47.
Dilek Hakkani-T ¨ur, Giuseppe Riccardi, and Allen
Gorin. 2002. Active learning for automatic speech
recognition. In Acoustics, Speech, and Signal Processing (ICASSP), 2002 IEEE International Conference on . IEEE, volume 4, pages IV–3904.
Matthew Henderson, Blaise Thomson, and Jason
Williams. 2013. Dialog state tracking challenge 2 &
3.http://camdial.org/ ˜mh521/dstc/ .
John E. Hopcroft, Rajeev Motwani, and Jeffrey D. Ullman. 2006. Introduction to Automata Theory, Languages, and Computation (3rd Edition) . AddisonWesley Longman Publishing Co., Inc., Boston, MA,
USA.
Bing Liu, Gokhan Tur, Dilek Hakkani-Tur, Pararth
Shah, and Larry Heck. 2017. End-to-end optimization of task-oriented dialogue model with deep reinforcement learning. In NIPS Conversational AI
Workshop .
Alexander H Miller, Will Feng, Adam Fisch, Jiasen Lu,
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
29
|
Workshop .
Alexander H Miller, Will Feng, Adam Fisch, Jiasen Lu,
Dhruv Batra, Antoine Bordes, Devi Parikh, and Jason Weston. 2017. Parlai: A dialog research software platform. arXiv preprint arXiv:1705.06476 .
Neha Nayak, Dilek Hakkani-Tur, Marilyn Walker, and
Larry Heck. 2017. To plan or not to plan? discourse planning in slot-value informed sequence to
sequence models for language generation. In Proc.
of Interspeech .
Abhinav Rastogi, Dilek Hakkani-Tur, and Larry Heck.
2017. Scalable multi-domain dialogue state tracking. In Proc. of IEEE ASRU .
Jost Schatzmann, Blaise Thomson, Karl Weilhammer,
Hui Ye, and Steve Young. 2007. Agenda-based user
simulation for bootstrapping a pomdp dialogue system. In Human Language Technologies 2007: The
Conference of the North American Chapter of the
Association for Computational Linguistics; Companion Volume, Short Papers . Association for Computational Linguistics, pages 149–152.
Pararth Shah, Dilek Hakkani-T ¨ur, and Larry Heck.
2016. Interactive reinforcement learning for taskoriented dialogue management. In NIPS Deep
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
30
|
Pararth Shah, Dilek Hakkani-T ¨ur, and Larry Heck.
2016. Interactive reinforcement learning for taskoriented dialogue management. In NIPS Deep
Learning for Action and Interaction Workshop .
David Silver, Aja Huang, Chris J Maddison, Arthur
Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. 2016. Mastering
the game of go with deep neural networks and tree
search. Nature 529(7587):484–489.David Silver, Thomas Hubert, Julian Schrittwieser,
Ioannis Antonoglou, Matthew Lai, Arthur Guez,
Marc Lanctot, Laurent Sifre, Dharshan Kumaran,
Thore Graepel, et al. 2017. Mastering chess
and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint
arXiv:1712.01815 .
PH Su, M Ga ˇsi´c, N Mrk ˇsi´c, L Rojas-Barahona, S Ultes,
D Vandyke, TH Wen, and S Young. 2016. Online active reward learning for policy optimisation in
spoken dialogue systems. In 54th Annual Meeting of
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
31
|
D Vandyke, TH Wen, and S Young. 2016. Online active reward learning for policy optimisation in
spoken dialogue systems. In 54th Annual Meeting of
the Association for Computational Linguistics, ACL
2016-Long Papers . volume 4, pages 2431–2441.
Gokhan Tur, Robert E Schapire, and Dilek HakkaniTur. 2003. Active learning for spoken language
understanding. In Acoustics, Speech, and Signal
Processing, 2003. Proceedings.(ICASSP’03). 2003
IEEE International Conference on . IEEE, volume 1,
pages I–I.
Stefan Ultes, Lina M Rojas Barahona, Pei-Hao Su,
David Vandyke, Dongho Kim, Inigo Casanueva,
Paweł Budzianowski, Nikola Mrk ˇsi´c, Tsung-Hsien
Wen, Milica Gasic, et al. 2017. Pydial: A multidomain statistical dialogue system toolkit. Proceedings of ACL 2017, System Demonstrations pages 73–
78.
Yushi Wang, Jonathan Berant, Percy Liang, et al. 2015.
Building a semantic parser overnight. ACL .
Tsung-Hsien Wen, David Vandyke, Nikola Mrksic,
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
32
|
Building a semantic parser overnight. ACL .
Tsung-Hsien Wen, David Vandyke, Nikola Mrksic,
Milica Gasic, Lina M Rojas-Barahona, Pei-Hao Su,
Stefan Ultes, and Steve Young. 2016. A networkbased end-to-end trainable task-oriented dialogue
system. ACL .
Victor Zhong, Caiming Xiong, and Richard Socher.
2017. Seq2sql: Generating structured queries
from natural language using reinforcement learning.
arXiv preprint arXiv:1709.00103 .
A Supplemental Material
Table 4 lists the dialogue acts used in our setup.
The dialogue acts are based on the Cambridge dialogue act set. Table 5 presents a full dialogue outline and corresponding paraphrase for a dialogue
spanning two interdependent tasks, where the user
wants to first buy movie tickets and then reserve a
restaurant table for dinner after the movie.
Figure 3 presents the interface shown to crowd
workers for the dialogue rewrite task, and includes
a sample dialogue outline (consisting of template
utterances) and its paraphrase into natural language. Figure 4 presents the interface shown to
crowd workers for evaluating the quality of dialogues collected with M2M.
Table 4: List of dialogue acts.
Dialogue Act Speaker Description
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
33
|
crowd workers for evaluating the quality of dialogues collected with M2M.
Table 4: List of dialogue acts.
Dialogue Act Speaker Description
GREETING User/System Greet the other speaker
INFORM User/System Inform a slot value
CONFIRM User/System Ask the other speaker to confirm a given slot value
REQUEST User/System Ask for the value of a slot
REQUEST ALTS User Ask for more alternatives
OFFER System Offer a database entity to the user
SELECT System Offer more than one database entity to the user
AFFIRM User/System Agree to something said by the other speaker
NEGATE User/System Disagree to something said by the other speaker
NOTIFY SUCCESS System Notify the user of a successful event, e.g. a booking is complete
NOTIFY FAILURE System Notify the user of a failure event, e.g. a booking isn’t available
THANK YOU User/System Thank the other speaker
GOOD BYE User/System Say goodbye to the other speaker
CANT UNDERSTAND User/System Tell the other speaker that their utterance was not understood
OTHER User Unknown utterance type
Table 5: Sample multi-domain dialogue outline and paraphrase.
Outline Paraphrase
Annotation ( ai) Template utterances ( ti) NL utterances ( ui)
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
1801.04871
|
34
|
Table 5: Sample multi-domain dialogue outline and paraphrase.
Outline Paraphrase
Annotation ( ai) Template utterances ( ti) NL utterances ( ui)
S: greeting() Greeting. Hi, how can I help you?
U: inform(intent=book movie,
name=Inside Out, date=tomorrow,
num tickets=2)Book movie with name is
Inside Out and date is tomorrow
and num tickets is 2.I want to buy 2 tickets for Inside
Out for tomorrow.
S: ack() request(time) OK. Provide time.Alright. What time would you like
to see the movie?
U: inform(time=evening) Time is evening.Anytime during the evening works
for me.
S: offer(theatre=Cinemark 16,
time=6pm)Offer theatre is Cinemark 16 and
time is 6pm.How about the 6pm show at
Cinemark 16?
U: affirm() Agree. That sounds good.
S: notify success() Reservation confirmed. Your tickets have been booked!
U: inform(intent=find restaurant,
meal=dinner, location=near the
theatre)Find restaurant with meal is dinner
and location is near the theatre.I want to get dinner at a restaurant
|
1801.04871
|
Building a Conversational Agent Overnight with Dialogue Self-Play
|
We propose Machines Talking To Machines (M2M), a framework combining
automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents
for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with
just a task schema and an API client from the dialogue system developer, but it
is also customizable to cater to task-specific interactions. Compared to the
Wizard-of-Oz approach for data collection, M2M achieves greater diversity and
coverage of salient dialogue flows while maintaining the naturalness of
individual utterances. In the first phase, a simulated user bot and a
domain-agnostic system bot converse to exhaustively generate dialogue
"outlines", i.e. sequences of template utterances and their semantic parses. In
the second phase, crowd workers provide contextual rewrites of the dialogues to
make the utterances more natural while preserving their meaning. The entire
process can finish within a few hours. We propose a new corpus of 3,000
dialogues spanning 2 domains collected with M2M, and present comparisons with
popular dialogue datasets on the quality and diversity of the surface forms and
dialogue flows.
|
http://arxiv.org/pdf/1801.04871
|
[
"Pararth Shah",
"Dilek Hakkani-Tür",
"Gokhan Tür",
"Abhinav Rastogi",
"Ankur Bapna",
"Neha Nayak",
"Larry Heck"
] |
[
"cs.AI",
"cs.CL"
] |
11 pages, 4 figures
| null |
cs.AI
|
20180115
|
20180115
|
[
{
"id": "1801.04871"
},
{
"id": "1709.00103"
},
{
"id": "1712.01815"
},
{
"id": "1705.06476"
},
{
"id": "1704.00057"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.