The article presents Soft Convex Quantization (SCQ) as a solution to the challenges of Vector Quantization (VQ) in deep learning. VQ is a technique used for extracting discrete latent representations and is used in various applications, including image and speech generation. However, VQ faces challenges like codebook collapse, non-differentiability, and lossy compression. SCQ is proposed as a direct substitute for VQ. It operates like a differentiable convex optimization layer, solving for the optimal convex combination of codebook vectors. The study shows that SCQ autoencoder models significantly outperform VQ-based architectures, providing better image reconstruction and codebook usage.
Publication date: 4 Oct 2023
Project Page: https://arxiv.org/abs/2310.03004
Paper: https://arxiv.org/pdf/2310.03004
