Disentangling Variational Autoencoders

TL;DR

This paper investigates how to achieve disentangled latent spaces in variational autoencoders, enabling more controllable and interpretable data generation, by analyzing different VAE models trained on handwritten digit images.

Contribution

It provides an experimental analysis of latent space disentanglement in VAEs, identifying factors that improve interpretability and control over generated data.

Findings

Latent dimensions can be aligned with visual properties like line weight, tilt, and width.

Increasing KL divergence contribution improves disentanglement.

Conditioning on input class enhances latent space interpretability.

Abstract

A variational autoencoder (VAE) is a probabilistic machine learning framework for posterior inference that projects an input set of high-dimensional data to a lower-dimensional, latent space. The latent space learned with a VAE offers exciting opportunities to develop new data-driven design processes in creative disciplines, in particular, to automate the generation of multiple novel designs that are aesthetically reminiscent of the input data but that were unseen during training. However, the learned latent space is typically disorganized and entangled: traversing the latent space along a single dimension does not result in changes to single visual attributes of the data. The lack of latent structure impedes designers from deliberately controlling the visual attributes of new designs generated from the latent space. This paper presents an experimental study that investigates latent space disentanglement. We implement three different VAE models from the literature and train them on a publicly available dataset of 60,000 images of hand-written digits. We perform a sensitivity analysis to find a small number of latent dimensions necessary to maximize a lower bound to the log marginal likelihood of the data. Furthermore, we investigate the trade-offs between the quality of the reconstruction of the decoded images and the level of disentanglement of the latent space. We are able to automatically align three latent dimensions with three interpretable visual properties of the digits: line weight, tilt and width. Our experiments suggest that i) increasing the contribution of the Kullback-Leibler divergence between the prior over the latents and the variational distribution to the evidence lower bound, and ii) conditioning input image class enhances the learning of a disentangled latent space with a VAE.