SimVAE: Simulator-Assisted Training forInterpretable Generative Models

TL;DR

SimVAE introduces a two-step training method for VAEs that uses a simulator to create a disentangled, interpretable latent space, improving training stability and applicability across scientific domains.

Contribution

The paper proposes a novel simulator-assisted training approach for VAEs that decouples encoder and decoder training, enhancing interpretability and training efficiency.

Findings

Produces a disentangled, interpretable latent space

Applicable to circuit design, graphics de-rendering, and natural sciences

Bypasses common training difficulties in VAEs and GANs

Abstract

This paper presents a simulator-assisted training method (SimVAE) for variational autoencoders (VAE) that leads to a disentangled and interpretable latent space. Training SimVAE is a two-step process in which first a deep generator network(decoder) is trained to approximate the simulator. During this step, the simulator acts as the data source or as a teacher network. Then an inference network (encoder)is trained to invert the decoder. As such, upon complete training, the encoder represents an approximately inverted simulator. By decoupling the training of the encoder and decoder we bypass some of the difficulties that arise in training generative models such as VAEs and generative adversarial networks (GANs). We show applications of our approach in a variety of domains such as circuit design, graphics de-rendering and other natural science problems that involve inference via simulation.