Kyle Roth

In this paper they show that we can achieve exponential performance scaling over dataset size, when the samples added are pruned to be only the best examples. This beats power law scaling in a big way. There is still no free lunch, in some sense, because in most cases it will become progressively harder to add new useful samples as the dataset gets bigger. But this is a big deal for computation, because it means that the number of samples in the dataset is not nearly as important as the coverage and quality that the dataset provides.

I think it’s valuable to be working in the open whenever possible, so I’m going to keep my research notes here. These notes will hopefully be full of good (and bad) ideas, so if someone borrows a good idea and publishes on it, that’s great! This post contains my research notes as I try to understand how model scaling affects worst-group performance. This started as a group project in the neural scaling laws course at Mila in winter 2022.

This was a paper I presented about in Bang Liu’s research group meeting on 2022-04-11. You can view the slides I used here.

The unsurprising result here is that few-shot performance scales predictably with pre-training dataset size under traditional fine-tuning, matching network, and prototypical network approaches. The interesting result is that the exponents of these three approaches were substantially different (see Table 1 in the paper), which says to me that the few-shot inference approach matters a lot. The surprising result was that while more training on the “non-natural” Omniglot dataset did not improve few-shot accuracy on other datasets, training on “natural” datasets did improve accuracy on few-shot Omniglot.

These authors define robust error as the least upper bound on the expected loss over a family of environmental settings (including dataset, model architecture, learning algorithm, etc.): \[\sup_{e\in\mathcal F}\mathbb E_{\omega\in P^e}\left[\ell(\phi,\omega)\right]\] The fact that this is an upper bound and not an average is very important and is what makes this work unique from previous work in this direction. Indeed, what we should be concerned about is not how poorly a model performs on the average sample but on the worst-case sample.