Semi-Amortized Variational Autoencoders

TL;DR

This paper introduces a semi-amortized variational autoencoder that combines amortized inference with local stochastic variational inference, improving generative modeling for text and images by avoiding posterior collapse.

Contribution

It proposes a hybrid inference method that refines amortized variational parameters with differentiable local SVI, enabling end-to-end training of richer models.

Findings

Outperforms autoregressive and variational baselines on text datasets

Reduces posterior collapse in training VAEs for text generation

Achieves better generative quality on image datasets

Abstract

Amortized variational inference (AVI) replaces instance-specific local inference with a global inference network. While AVI has enabled efficient training of deep generative models such as variational autoencoders (VAE), recent empirical work suggests that inference networks can produce suboptimal variational parameters. We propose a hybrid approach, to use AVI to initialize the variational parameters and run stochastic variational inference (SVI) to refine them. Crucially, the local SVI procedure is itself differentiable, so the inference network and generative model can be trained end-to-end with gradient-based optimization. This semi-amortized approach enables the use of rich generative models without experiencing the posterior-collapse phenomenon common in training VAEs for problems like text generation. Experiments show this approach outperforms strong autoregressive and variational baselines on standard text and image datasets.