Preventing Posterior Collapse with delta-VAEs
This addresses a key bottleneck in variational autoencoders for researchers and practitioners in generative modeling, offering a method to prevent posterior collapse while maintaining model performance, though it is an incremental improvement over existing techniques.
The paper tackles the problem of posterior collapse in latent variable generative models by proposing δ-VAEs, which constrain the variational posterior to maintain a minimum distance from the prior, enabling the use of powerful decoders without sacrificing latent information. This approach achieved state-of-the-art log-likelihood on CIFAR-10 and ImageNet 32×32, improving sample quality and representation learning.
Due to the phenomenon of "posterior collapse," current latent variable generative models pose a challenging design choice that either weakens the capacity of the decoder or requires augmenting the objective so it does not only maximize the likelihood of the data. In this paper, we propose an alternative that utilizes the most powerful generative models as decoders, whilst optimising the variational lower bound all while ensuring that the latent variables preserve and encode useful information. Our proposed $δ$-VAEs achieve this by constraining the variational family for the posterior to have a minimum distance to the prior. For sequential latent variable models, our approach resembles the classic representation learning approach of slow feature analysis. We demonstrate the efficacy of our approach at modeling text on LM1B and modeling images: learning representations, improving sample quality, and achieving state of the art log-likelihood on CIFAR-10 and ImageNet $32\times 32$.