Resetting a fixed broken ELBO
This provides a probabilistically sound alternative to beta-VAEs for improving latent space disentanglement in generative modeling.
The paper addresses the violation of probability laws in beta-VAEs by introducing a Renyi entropy VAE with a new RELBO, which includes a parameterized regularizer term evaluated analytically via SVD.
Variational autoencoders (VAEs) are one class of generative probabilistic latent-variable models designed for inference based on known data. They balance reconstruction and regularizer terms. A variational approximation produces an evidence lower bound (ELBO). Multiplying the regularizer term by beta provides a beta-VAE/ELBO, improving disentanglement of the latent space. However, any beta value different than unity violates the laws of conditional probability. To provide a similarly-parameterized VAE, we develop a Renyi (versus Shannon) entropy VAE, and a variational approximation RELBO that introduces a similar parameter. The Renyi VAE has an additional Renyi regularizer-like term with a conditional distribution that is not learned. The term is evaluated essentially analytically using a Singular Value Decomposition method.