Towards Visually Explaining Variational Autoencoders

TL;DR

This paper introduces a novel gradient-based attention method to visually explain variational autoencoders, enabling anomaly localization and improved latent space disentanglement, bridging a gap in interpretability for generative models.

Contribution

It presents the first technique for visual explanations of VAEs using gradient-based attention, applicable to anomaly detection and model training enhancement.

Findings

Achieves state-of-the-art anomaly localization on MVTec-AD.

Improves latent space disentanglement on Dsprites.

Demonstrates the utility of attention maps beyond prediction explanation.

Abstract

Recent advances in Convolutional Neural Network (CNN) model interpretability have led to impressive progress in visualizing and understanding model predictions. In particular, gradient-based visual attention methods have driven much recent effort in using visual attention maps as a means for visual explanations. A key problem, however, is these methods are designed for classification and categorization tasks, and their extension to explaining generative models, e.g. variational autoencoders (VAE) is not trivial. In this work, we take a step towards bridging this crucial gap, proposing the first technique to visually explain VAEs by means of gradient-based attention. We present methods to generate visual attention from the learned latent space, and also demonstrate such attention explanations serve more than just explaining VAE predictions. We show how these attention maps can be used to localize anomalies in images, demonstrating state-of-the-art performance on the MVTec-AD dataset. We also show how they can be infused into model training, helping bootstrap the VAE into learning improved latent space disentanglement, demonstrated on the Dsprites dataset.