Most people I know who publish in NIPS/ICLR train their models using between 1-4 GPUs under their desk or 8 GPUs in a server. I would argue that these folks make competitive models and "are successful in their business" with only 1-8 GPUs. While having access to hundreds of GPUs helps you explore model space more quickly, it's not a hard pre-requisite to success in Deep Learning.

Yup, state of the art doesn't require GPU farms, just a couple in a consumer desktop will do.

Lookup AmoebaNet-D if you think you don't need a GPU farm these days. Those were times when 1-2 GPUs were enough...

You know, there is more to deep learning research than meta-learning/architecture exploration. Sure you can explore the hyper-parameter space faster with 500 GPUs and get yet again a 0.05% better test score on ImageNet (or more I don't actually know), but there are other ways to do something meaningful in DL without using such compute power.

That's a fair point and I agree. It's just sometimes difficult to beat automated exploration; as a standard company you probably don't have access to top-end researchers/practitioners, just average ones, and those might get a significant boost by trading smartness for brute force and run many models in parallel in evolutionary fashion.

When you e.g. look at how Google's internal system constructs loss functions and how many combinations do they try, one has to have an unexpected idea to beat their results, and that idea can be usually quickly incorporated into their platform, raising the bar for individual researchers. At Facebook they basically press a few buttons and select a few checkboxes, then wait until best model is selected, leading to frustration among researchers.

It's just an indicator that extensive improvement is possible. But as with adding more and more shader cores you can get your FPS gainz to the point of Ray-tracing appears around the corner and now you need to add different kind of cores.

Same process for research. You suppose to find some insights on how to do one thing or another, find the direction of search, eventually there would be hardware to fully explore that direction. Then you move on to a different direction. Rinse-repeat

We used to call this model fishing and regard models that came from places that did it with fear and suspicion (due to sudden failure and poor performance in production)

What has changed that people think this is a wise approach?

They don't suddenly fail any more, and performance in production is fine... This is a rather empirical science right now.

If forced to speculate, I'd say far larger datasets are part of the answer. If you can afford to hold back 30%+ of your data for (multi-tranche) verification the difference between testing and production becomes a problem for the philosophers.

Hmm this isn't my experience. My experience is that highly complex models tend to reflect accurately the ground truth in the current dataset, but not the domain theory that generated it, so when things change (like, prevailing weather conditions, political alignments, interest rates, data usage) which move the distribution (but not the domain theory) they do fail. The question is : are you trading poorer compression for more fidelity or are you learning the system that generates the data?

AutoML and automated architecture/hyperparameter search used by FAANG, producing things like AmoebaNet-D, that beat all previous state-of-art results.