Discouraging posterior collapse in hierarchical Variational Autoencoders using context
This addresses a persistent issue in generative modeling for researchers and practitioners, but is incremental as it builds on existing hierarchical VAE frameworks.
The paper tackles the problem of posterior collapse in hierarchical Variational Autoencoders, showing it persists despite common assumptions, and proposes a modification using a Discrete Cosine Transform context to improve latent space utilization without harming generative performance.
Hierarchical Variational Autoencoders (VAEs) are among the most popular likelihood-based generative models. There is a consensus that the top-down hierarchical VAEs allow effective learning of deep latent structures and avoid problems like posterior collapse. Here, we show that this is not necessarily the case, and the problem of collapsing posteriors remains. To discourage this issue, we propose a deep hierarchical VAE with a context on top. Specifically, we use a Discrete Cosine Transform to obtain the last latent variable. In a series of experiments, we observe that the proposed modification allows us to achieve better utilization of the latent space and does not harm the model's generative abilities.