Training Neural Networks for Modularity aids Interpretability

Abstract

An approach to improve network interpretability is via clusterability, i.e., splitting a model into disjoint clusters that can be studied independently. We find pretrained models to be highly unclusterable and thus train models to be more modular using an ``enmeshment loss'' function that encourages the formation of non-interacting clusters. Using automated interpretability measures, we show that our method finds clusters that learn different, disjoint, and smaller circuits for CIFAR-10 labels. Our approach provides a promising direction for making neural networks easier to interpret.

Some of the interpretations of the results in this paper were incorrect, and we found on further experiments that the techniques did not scale well - we have the corrected results in a different submission (due to a significant change in both content and authors it needed to be a new submission). Please check https://arxiv.org/abs/2502.02470 for the updated paper.