Important paper from Google on large batch optimization. They do impressively careful experiments measuring # iterations needed to achieve target validation error at various batch sizes. The main "surprise" is the lack of surprises. [thread]

The paper is a good example of lots of elements of good experimental design. They validate their metric by showing lots of variants give consistent results. They tune hyperparamters separately for each condition, check that optimum isn't at the endpoints, and measure sensitivity.
They have separate experiments where the hold fixed # iterations and # epochs, which (as they explain) measure very different things. They avoid confounds, such as batch norm's artificial dependence between batch size and regularization strength.
When the experiments are done carefully enough, the results are remarkably consistent between different datasets and architectures. Qualitatively, MNIST behaves just like ImageNet.
Importantly, they don't find any evidence for a "sharp/flat optima" effect whereby better optimization leads to worse final results. They have a good discussion of experimental artifacts/confounds in past papers where such effects were reported.
The time-to-target-validation is explained purely by optimization considerations. There's a regime where variance dominates, and you get linear speedups w/ batch size. Then there's a regime where curvature dominates and larger batches don't help. As theory would predict.
Incidentally, this paper must have been absurdly expensive, even by Google's standards. Doing careful empirical work on optimizers requires many, many runs of the algorithm. (I think surprising phenomena on ImageNet are often due to the difficulty of running proper experiments.)

More from Machine learning

Really enjoyed digging into recent innovations in the football analytics industry.

>10 hours of interviews for this w/ a dozen or so of top firms in the game. Really grateful to everyone who gave up time & insights, even those that didnt make final cut 🙇‍♂️

For avoidance of doubt, leading tracking analytics firms are now well beyond voronoi diagrams, using more granular measures to assess control and value of space.

This @JaviOnData & @LukeBornn paper from 2018 referenced in the piece demonstrates one method

Bit of this that I nerded out on the most is "ghosting" — technique used by @counterattack9 & co @stats_insights, among others.

Deep learning models predict how specific players — operating w/in specific setups — will move & execute actions. A paper here:

So many use-cases:
1/ Quickly & automatically spot situations where opponent's defence is abnormally vulnerable. Drill those to death in training.
2/ Swap target player B in for current player A, and simulate. How does target player strengthen/weaken team? In specific situations?

You May Also Like