Exposition and Interpretation of the Topology of Neural Networks

TL;DR

This paper applies topological data analysis to CNNs, revealing simple global structures in weights, how these structures evolve during training, and their correlation with generalization ability, offering insights into interpretability and performance improvement.

Contribution

It introduces a topological data model for CNN weights, demonstrating how topological structures relate to training dynamics and generalization, and shows how this information can enhance network performance.

Findings

Weights learn simple global topological structures.

Topological structures change during training.

Topological features correlate with generalization ability.

Abstract

Convolutional neural networks (CNN's) are powerful and widely used tools. However, their interpretability is far from ideal. One such shortcoming is the difficulty of deducing a network's ability to generalize to unseen data. We use topological data analysis to show that the information encoded in the weights of a CNN can be organized in terms of a topological data model and demonstrate how such information can be interpreted and utilized. We show that the weights of convolutional layers at depths from 1 through 13 learn simple global structures. We also demonstrate the change of the simple structures over the course of training. In particular, we define and analyze the spaces of spatial filters of convolutional layers and show the recurrence, among all networks, depths, and during training, of a simple circle consisting of rotating edges, as well as a less recurring unanticipated complex circle that combines lines, edges, and non-linear patterns. We also demonstrate that topological structure correlates with a network's ability to generalize to unseen data and that topological information can be used to improve a network's performance. We train over a thousand CNN's on MNIST, CIFAR-10, SVHN, and ImageNet.