SQ-VAE: Variational Bayes on Discrete Representation with Self-annealed Stochastic Quantization
Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi
Abstract
One noted issue of vector-quantized variational autoencoder (VQ-VAE) is that the learned discrete representation uses only a fraction of the full capacity of the codebook, also known as codebook collapse. We hypothesize that the training scheme of VQ-VAE, which involves some carefully designed heuristics, underlies this issue. In this paper, we propose a new training scheme that extends the standard VAE via novel stochastic dequantization and quantization, called stochastically quantized variational autoencoder (SQ-VAE). In SQ-VAE, we observe a trend that the quantization is stochastic at the initial stage of the training but gradually converges toward a deterministic quantization, which we call self-annealing. Our experiments show that SQ-VAE improves codebook utilization without using common heuristics. Furthermore, we empirically show that SQ-VAE is superior to VAE and VQ-VAE in vision- and speech-related tasks.
BibTeX
@InProceedings{pmlr-v162-takida22a,
title = {{SQ}-{VAE}: Variational {B}ayes on Discrete Representation with Self-annealed Stochastic Quantization},
author = {Takida, Yuhta and Shibuya, Takashi and Liao, Weihsiang and Lai, Chieh-Hsin and Ohmura, Junki and Uesaka, Toshimitsu and Murata, Naoki and Takahashi, Shusuke and Kumakura, Toshiyuki and Mitsufuji, Yuki},
booktitle = {Proceedings of the 39th International Conference on Machine Learning},
pages = {20987--21012},
year = {2022},
editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
volume = {162},
series = {Proceedings of Machine Learning Research},
month = {17--23 Jul},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v162/takida22a/takida22a.pdf},
url = {https://proceedings.mlr.press/v162/takida22a.html},
abstract = {One noted issue of vector-quantized variational autoencoder (VQ-VAE) is that the learned discrete representation uses only a fraction of the full capacity of the codebook, also known as codebook collapse. We hypothesize that the training scheme of VQ-VAE, which involves some carefully designed heuristics, underlies this issue. In this paper, we propose a new training scheme that extends the standard VAE via novel stochastic dequantization and quantization, called stochastically quantized variational autoencoder (SQ-VAE). In SQ-VAE, we observe a trend that the quantization is stochastic at the initial stage of the training but gradually converges toward a deterministic quantization, which we call self-annealing. Our experiments show that SQ-VAE improves codebook utilization without using common heuristics. Furthermore, we empirically show that SQ-VAE is superior to VAE and VQ-VAE in vision- and speech-related tasks.}
}