[link]
Summary: An odd thing about machine learning these days is how far you can get in a line of research while only ever testing your method on image classification and image datasets in general. This leads one occasionally to wonder whether a given phenomenon or advance is a discovery of the field generally, or whether it's just a fact about the informatics and learning dynamics inherent in image data. This paper, part of a set of recent papers released by Facebook centering around the Lottery Ticket Hypothesis, exists in the noble tradition of "lets try <thing> on some nonimage datasets, and see if it still works". This can feel a bit silly in the cases where the ideas or approaches do transfer, but I think it's still an important impulse for the field to have, lest we become too captured by ImageNet and its various descendants. This paper test the Lottery Ticket Hypothesis  the idea that there are a small subset of weights in a trained network whose lucky initializations promoted learning, such that if you reset those weights to their initializations and train only them you get comparable or nearcomparable performance to the full network  on reinforcement learning and NLP datasets. In particular, within RL, they tested on both simple continuous control (where the observation state is a vector of meaningful numbers) and Atari from pixels (where the observation is a full frompixels image). In NLP, they trained on language modeling and translation, with both a LSTM and a Transformer respectively. (Prior work had found that Transformers didn't exhibit lottery ticket like phenomenon, but this paper found a circumstance where they appear to. ) Some high level interesting results: https://i.imgur.com/kd03bQ4.png https://i.imgur.com/rZTH7FJ.png  So as to not bury the lede: by and large, "winning" tickets retrained at their original initializations outperform random initializations of the same size and configuration on both NLP and Reinforcement Learning problems  There is wide variability in how much pruning in general (a necessary prerequisite operation) impacts reinforcement learning. On some games, pruning at all crashes performance, on others, it actually improves it. This leads to some inherent variability in results https://i.imgur.com/4o71XPt.png  One thing that prior researchers in this area have found is that pruning weights all at once at the end of training tends to crash performance for complex models, and that in order to find pruned models that have Lottery Ticketesque highperforming properties, you need to do "iterative pruning". This works by training a model for a period, then pruning some proportion of weights, then training again from the beginning, and then pruning again, and so on, until you prune down to the full percentage you want to prune. The idea is that this lets the model adapt gradually to a drop in weights, and "train around" the capacity reduction, rather than it just happening all at once. In this paper, the authors find that this is strongly necessary for Lottery Tickets to be found for either Transformers or many RL problems. On a surface level, this makes sense, since Reinforcement Learning is a notoriously tricky and nonstationary learning problem, and Transformers are complex models with lots of parameters, and so dramatically reducing parameters can handicap the model. A weird wrinkle, though, is that the authors find that lottery tickets found without iterative pruning actually perform worse than "random tickets" (i.e. initialized subnetworks with random topology and weights). This is pretty interesting, since it implies that the topology and weights you get if you prune all at once are actually counterproductive to learning. I don't have a good intuition as to why, but would love to hear if anyone does. https://i.imgur.com/9LnJe6j.png  For the Transformer specifically, there was an interesting divergence in the impact of weight pruning between the weights of the embedding matrix and the weights of the rest of the network machinery. If you include embeddings in the set of weights being pruned, there's essentially no difference in performance between random and winning tickets, whereas if you exclude them, winning tickets exhibit the more typical pattern of outperforming random ones. This implies that whatever phenomenon that makes winning tickets better is more strongly (or perhaps only) present in weights for feature calculation on top of embeddings, and not very present for the embeddings themselves
Your comment:
