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]

https://t.co/7QIx5CFdfJ

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.

1 There's a chasm between an NLP technology that works well in the research lab and something that works for applications that real people use. This was eye-opening when I started my career, and every time I talk to an NLP engineer at @textio, it continues to strike me even now.

2 Research conditions are theoretical and/or idealized. A huge problem for so-called NLP or AI startups with highly credentialed academic founders is that they bring limited knowledge of what it takes to build real products outside the lab.

3 A product is ultimately a thing that people pay for - not just cool technology or user experience. But I’m not even talking about knowledge gaps in go-to-market work. I'm talking purely technical gaps: how you go from science project to performant + delightful user experience.

4 Most commoditized NLP packages solve well-understood problems in standard ways that sacrifice either precision or performance. In a research lab, this is not usually a hard trade-off; in general, no one is using what you make, so performance is less important than precision.

5 In software, when you’re making something for real people to use, these tradeoffs are a big deal. Especially if you’re asking those people to pay for what you’ve made (can’t get away from that pesky GTM thinking). They expect quality, which includes precision AND performance.