SimVAE: Simulator-Assisted Training forInterpretable Generative Models
This method addresses the problem of interpretability in generative models for researchers and practitioners in fields involving simulation-based inference, though it appears incremental as it builds on existing VAE and simulator frameworks.
The paper tackles the challenge of training interpretable generative models by introducing SimVAE, a two-step simulator-assisted training method for variational autoencoders that results in a disentangled and interpretable latent space. It demonstrates applications in domains like circuit design and graphics de-rendering.
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.