Reproduce “Regularized Residual Quantization: a multi-layer sparse dictionary learning approach” on OpenResearch

About this paper

Sohrab Ferdowsi, Slava Voloshynovskiy, Dimche Kostadinov

Researchers at the University of Geneva introduce Regularized Residual Quantization (RRQ), a multi-layer sparse dictionary learning approach that uses variance-based regularization to overcome overfitting in traditional residual quantization. The system demonstrates a reduced train-test distortion gap and successfully learns multi-layer codebooks for high-dimensional image processing, enabling enhanced super-resolution.

Key insights

• Insights from rate-distortion theory can effectively regularize codebook variances in quantization, significantly reducing overfitting and improving generalization to unseen data.
• Variance-based regularization naturally induces sparsity in learned codewords by assigning zero rate to low-variance dimensions, leading to more efficient representations.
• A pre-processing strategy combining global 2D-DCT with local PCA effectively decorrelates high-dimensional image data, making it suitable for multi-layer residual quantization.

Summary by alphaXiv.