Reproduce “Convolutional Dictionary Learning: A Comparative Review and New Algorithms” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cristina Garcia-Cardona, Brendt Wohlberg
Researchers at Los Alamos National Laboratory provide the first comprehensive performance comparison of Convolutional Dictionary Learning (CDL) algorithms, identifying the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) as the top performer for serial processing and an ADMM Consensus approach for parallel processing. The paper also introduces new high-performing algorithms and provides practical heuristics for parameter selection, along with open-source implementations.
Key insights
- Optimal algorithm parameters for CDL are stable across similar training sets and scale consistently (constant or linear with the number of images K), enabling simple and reliable heuristic parameter selection.
- The 3D method by Šorel and Šroubek is a specific instance of the ADMM Consensus framework in the DFT domain, unifying previous distinct approaches.
- Training set performance of the CDL functional serves as a reliable predictor for the quality and generalization ability of the learned dictionary on unseen data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.