Abstract

State-of-the-art neural network architectures continue to scale in size and deliver impressive results on unseen data points at the expense of poor interpretability. In the deep layers of these models we often encounter very high dimensional feature spaces, where constructing graphs from intermediate data representations can lead to the well-known curse of dimensionality. We propose a channel-wise graph construction method that works on lower dimensional subspaces and provides a new channel-based perspective that leads to better interpretability of the data and relationship between channels. In addition, we introduce a novel generalization estimate based on the proposed graph construction method with which we perform local polytope interpolation. We show its potential to replace the standard generalization estimate based on validation set performance to perform progressive channel-wise early stopping without requiring a validation set.