Autoencoder Node Saliency: Selecting Relevant Latent Representations

TL;DR

This paper introduces a supervised node saliency method for autoencoders that ranks hidden nodes based on their class-distribution properties, helping interpret learned representations.

Contribution

The paper proposes a novel SNS method that uses class distribution comparisons and entropy measures to identify important nodes in autoencoders, enhancing interpretability.

Findings

SNS effectively ranks nodes by relevance to class distinctions.

NED identifies nodes with strong classifying properties.

Method applied successfully to real datasets.

Abstract

The autoencoder is an artificial neural network model that learns hidden representations of unlabeled data. With a linear transfer function it is similar to the principal component analysis (PCA). While both methods use weight vectors for linear transformations, the autoencoder does not come with any indication similar to the eigenvalues in PCA that are paired with the eigenvectors. We propose a novel supervised node saliency (SNS) method that ranks the hidden nodes by comparing class distributions of latent representations against a fixed reference distribution. The latent representations of a hidden node can be described using a one-dimensional histogram. We apply normalized entropy difference (NED) to measure the "interestingness" of the histograms, and conclude a property for NED values to identify a good classifying node. By applying our methods to real data sets, we demonstrate the ability of SNS to explain what the trained autoencoders have learned.