Background

Two things determine how much compute it is useful to throw at training an AI system: first, economics (the supply and demand for compute, which determine when the cost becomes infeasible). Second, parallelizability.

To understand this concept, suppose you wanted to add up a list of 1000 numbers. If you could get 10 friends to come in and help you, you could add up 100 numbers each and finish the task much faster. The task is parallelizable. However, suppose you wanted to list all the numbers in the Fibonacci sequence (where each entry is the sum of the two preceding it). Then having 10 friends join you wouldn’t be as helpful, since to compute every new number in the sequence you need to know what the last number was. So you all couldn’t just race ahead at different speeds.

Similarly, in order for researchers to benefit from larger compute, they must find ways of parallelising their training, either by splitting the data and running many instances of the same model across many machines, or by splitting the model itself in a way which can be efficiently parallelized.

A recent OpenAI milestone release notes that:

Batch sizes of 8 thousand [GDG+17], 16 thousand [SKYL17], 32 thousand [YGG17, YZH+17, ASF17], and even 64 thousand [JSH+18] examples have been effectively employed to train ImageNet, and batch sizes of thousands have been effective for language models and generative models [OEGA18, PKYC18, BDS18].

This phenomenon is not confined to supervised learning: in reinforcement learning, batch sizes of over a million timesteps (with tens of thousands of environments running in parallel) have been used in a Dota-playing agent [BCD+18], and even in simple Atari environments batch sizes of several thousand timesteps have proved effective [AAG+18, HQB+18, SA18]

We now ask:

By Jan 1, 2020, what will be the maximum batch size used in training by a published AI system?

Resolution

A “published AI system” is a system that is the topic of a published research paper, pre-print or credible blogpost. In order to be admissible, the paper/blog post must give sufficient information to estimate training compute, within some error threshold.

Data

The current record is a batch size of ~8.4M observations, by OpenAI's 1v1 Dota bot.

We have yet to compile an outside of the historical growth in batch sizes. Contributions here are very welcome!

Acknowledgements

This question was suggested by Tamay in the question suggestion thread.