Interpreting Neural Networks through the Polytope Lens

TL;DR

This paper introduces the polytope lens, a new approach to interpret neural networks by analyzing the partitioning of activation space into polytopes, offering clearer insights than traditional neuron or direction-based methods.

Contribution

The paper proposes the polytope lens as a fundamental unit for neural network interpretability, addressing limitations of neuron and direction-based descriptions by focusing on activation space polytopes.

Findings

Polytopes identify monosemantic regions of activation space.

Density of polytope boundaries correlates with semantic boundaries.

Polytope analysis predicts neural network behavior effectively.

Abstract

Mechanistic interpretability aims to explain what a neural network has learned at a nuts-and-bolts level. What are the fundamental primitives of neural network representations? Previous mechanistic descriptions have used individual neurons or their linear combinations to understand the representations a network has learned. But there are clues that neurons and their linear combinations are not the correct fundamental units of description: directions cannot describe how neural networks use nonlinearities to structure their representations. Moreover, many instances of individual neurons and their combinations are polysemantic (i.e. they have multiple unrelated meanings). Polysemanticity makes interpreting the network in terms of neurons or directions challenging since we can no longer assign a specific feature to a neural unit. In order to find a basic unit of description that does not suffer from these problems, we zoom in beyond just directions to study the way that piecewise linear activation functions (such as ReLU) partition the activation space into numerous discrete polytopes. We call this perspective the polytope lens. The polytope lens makes concrete predictions about the behavior of neural networks, which we evaluate through experiments on both convolutional image classifiers and language models. Specifically, we show that polytopes can be used to identify monosemantic regions of activation space (while directions are not in general monosemantic) and that the density of polytope boundaries reflect semantic boundaries. We also outline a vision for what mechanistic interpretability might look like through the polytope lens.