Theory and Experiments on Vector Quantized Autoencoders
This work addresses the performance gap in discrete latent variable models for researchers in machine learning, offering incremental improvements in training efficiency and speed for tasks like image generation and translation.
The authors tackled the challenge of training discrete latent variable models by proposing an EM-inspired training technique for VQ-VAE, achieving better image generation on CIFAR-10 and enabling a non-autoregressive machine translation model that is 3.3 times faster at inference while nearly matching the accuracy of a strong baseline Transformer.
Deep neural networks with discrete latent variables offer the promise of better symbolic reasoning, and learning abstractions that are more useful to new tasks. There has been a surge in interest in discrete latent variable models, however, despite several recent improvements, the training of discrete latent variable models has remained challenging and their performance has mostly failed to match their continuous counterparts. Recent work on vector quantized autoencoders (VQ-VAE) has made substantial progress in this direction, with its perplexity almost matching that of a VAE on datasets such as CIFAR-10. In this work, we investigate an alternate training technique for VQ-VAE, inspired by its connection to the Expectation Maximization (EM) algorithm. Training the discrete bottleneck with EM helps us achieve better image generation results on CIFAR-10, and together with knowledge distillation, allows us to develop a non-autoregressive machine translation model whose accuracy almost matches a strong greedy autoregressive baseline Transformer, while being 3.3 times faster at inference.