Interpretable Neuron Structuring with Graph Spectral Regularization

TL;DR

This paper introduces Graph Spectral Regularization to improve interpretability of neural network hidden layers by encouraging smoothness in neuron activations, inspired by biological spatial organization, without significantly affecting performance.

Contribution

It proposes a novel regularization method that structures neuron activations using graph Laplacian penalties, enhancing interpretability and visualization capabilities.

Findings

Enhanced interpretability of neural networks through structured activations.

Effective visualization and clustering in biological and image datasets.

Maintained primary task performance with the regularization.

Abstract

While neural networks are powerful approximators used to classify or embed data into lower dimensional spaces, they are often regarded as black boxes with uninterpretable features. Here we propose Graph Spectral Regularization for making hidden layers more interpretable without significantly impacting performance on the primary task. Taking inspiration from spatial organization and localization of neuron activations in biological networks, we use a graph Laplacian penalty to structure the activations within a layer. This penalty encourages activations to be smooth either on a predetermined graph or on a feature-space graph learned from the data via co-activations of a hidden layer of the neural network. We show numerous uses for this additional structure including cluster indication and visualization in biological and image data sets.