<ul><li>Vector quantization (VQ) is a technique used to learn features with discrete codebook representations.</li><li>Existing hierarchical extensions of VQ-VAE suffer from the codebook/layer collapse issue, leading to degraded reconstruction accuracy.</li><li>To address this problem, a novel framework called HQ-VAE (hierarchically quantized variational autoencoder) is proposed.</li><li>HQ-VAE improves codebook usage and enhances reconstruction performance in image datasets and audio datasets.</li></ul>

Research on inner workings of Variational Bayes method part3(Statistical Machine Learning)

Discover more