Introduction

Last week's paper roundup (more or less by accident) focused mostly on path dependence of deep learning and the order of feature learning. Going forwards, I've decided to have an explicit focus for each week's roundup. This week's focus is on the structure/redundancy of trained models, as well as linear interpolations through parameter space.

I've also decided to publish each roundup on Monday morning (Edit: or try to, at any rate).

Papers

Residual Networks Behave Like Ensembles of Relatively Shallow Networks

In this work we propose a novel interpretation of residual networks showing that they can be seen as a collection of many paths of differing length. Moreover, residual networks seem to enable very deep networks by leveraging only the short paths during training. To support this observation, we rewrite residual networks as an explicit collection of paths. Unlike traditional models, paths through residual networks vary in length. Further, a lesion study reveals that these paths show ensemble-like behavior in the sense that they do not strongly depend on each other. Finally, and most surprising, most paths are shorter than one might expect, and only the short paths are needed during training, as longer paths do not contribute any gradient. For example, most of the gradient in a residual network with 110 layers comes from paths that are only 10-34 layers deep. Our results reveal one of the key characteristics that seem to enable the training of very deep networks: Residual networks avoid the vanishing gradient problem by introducing short paths which can carry gradient throughout the extent of very deep networks.

My opinion:

This paper suggests that neural nets are redundant by default, which gives some intuition for why it's often possible to prune large fractions of a network's parameters without much impact on the test performance, as well as the mechanism by which residual connections allow for training deeper networks: residual connections allow shallow nets to communicate directly with the input / output space, so they allow for deep nets to be built from ensembling shallow nets.

I think it also points away from neural nets implementing a Kolmogorov or circuit simplicity prior.

On the Effect of Dropping Layers of Pre-trained Transformer Models

Transformer-based NLP models are trained using hundreds of millions or even billions of

...

Quintin's Alignment Papers Roundup