Neural Representations Reveal Distinct Modes of Class Fitting in Residual Convolutional Networks

TL;DR

This paper uses probabilistic models to analyze how residual networks fit classes, revealing two distinct modes of class representation that relate to memorization and robustness, especially in deeper layers.

Contribution

It uncovers two different modes of class fitting in residual networks' deep layers using class-conditional density models, linking these modes to memorization and robustness.

Findings

Classes are fitted with two distinct distribution modes.

Deeper layers reveal these modes, not low-level features.

Representation structures correlate with memorization and robustness.

Abstract

We leverage probabilistic models of neural representations to investigate how residual networks fit classes. To this end, we estimate class-conditional density models for representations learned by deep ResNets. We then use these models to characterize distributions of representations across learned classes. Surprisingly, we find that classes in the investigated models are not fitted in an uniform way. On the contrary: we uncover two groups of classes that are fitted with markedly different distributions of representations. These distinct modes of class-fitting are evident only in the deeper layers of the investigated models, indicating that they are not related to low-level image features. We show that the uncovered structure in neural representations correlate with memorization of training examples and adversarial robustness. Finally, we compare class-conditional distributions of neural representations between memorized and typical examples. This allows us to uncover where in the network structure class labels arise for memorized and standard inputs.