In 2018, a group including many of the authors of this updated paper argued for a theory of deep neural network optimization that they called the “Lottery Ticket Hypothesis”. It framed itself as a solution to what was otherwise a confusing seeming-contradiction: that you could prune or compress trained networks to contain a small percentage of their trained weights without loss of performance, but also that if you tried to train a comparably small network (comparable to the post-training pruned network) from initialization, you wouldn’t be able to achieve similar performance. They showed that, at least for some set of tasks, you could in fact train a smaller network to equivalent performance, but only if you kept the same connectivity patterns as in the pruned network, and if you re-initialized those neurons to the same values they were initialized at during the initial training. These lucky initialization patterns are the lottery tickets being referred to in the eponymous hypothesis: small subnetworks of well-initialized weights that are set up to be able to train well. This paper assesses whether and under what conditions the LTH holds on larger problems, and does a bit of a meta-analysis over different alternate theories in this space. One such alternate theory, from Liu et al, proposes that, in fact, there is no value in re-initializing to the specific initial values, and that you can actually get away with random initialization if you keep the connectivity patterns of the pruned weights. The “At Scale” paper compares the two methods over a wide range of pruning percentages, and convincingly shows that while random initialization with the same connectivity can perform well up to 80% of the weights being removed, after 80%, the performance of the random initialization drops, whereas the performance of the “winning ticket” approach remains comparable with full network training up to 99% of the weights being pruned. This seems to provide support for the theory that there is value in re-initializing the weights to how they were, especially when you prune to very small subnetworks. https://i.imgur.com/9O2aAIT.png The core of the current paper focuses on a difficulty in the original LTH paper: that the procedure of iterative pruning (train, then prune some weights, then train again) wasn’t able to reliably find “winning tickets” for deep networks of the type needed to solve ImageNet or CIFAR. To be precise, re-initializing pruned networks to their original values did no better than initializing them randomly in these networks. In order to actually get these winning tickets to perform well, the original authors had to do a somewhat arcane process of of starting the learning rate very small and scaling it up, called “warmup”. Neither paper gave a clear intuition as to why this would be the case, but the updated paper found that they could avoid the need for this approach if, instead of re-initializing weights to their original value, they set them to the values they were at after some small number of iterations into training. They justify this by showing that performance under this new initialization is related to something they call “stability to pruning,” which measures how close the learned weights after re-initialization are to the original learned weights in the full model training. And, while the weights of deeper networks are unstable (by this metric) when first initialized, they become stable fairly early on. I was a little confused by this framing, since it seemed fairly tautological to me, since you’re using “how stably close are the weights to the original weights” as a way to explain “when can you recover performance comparable to original performance.” This was framed as being a mechanistic explanation of why you can see a lottery ticket phenomenon to some extent, but only if you do a “late reset” to several iterations after initialization, but it didn’t feel quite mechanistically satisfying enough to me. That said, I think this is overall an intriguing idea, and I’d love to see more papers discuss it. In particular, I’d love to read more qualitative analysis about whether there are any notable patterns shared by “winning tickets”.