Implementation and experiments for an AISTATS 2018 (theory) paper on covergence rates of convolutional sparse dictionary learning. The paper can be found here.
An early version of the full paper, presented at the 2017 Allerton Conference on Communication, Control, and Computing, is available here.
The figures in the paper can be generated by running the files experiment1.m, experiment2.m, experiment2_correlated.m, experiment3.m, and experiment4.m, respectively, in the reconstruction_synthetic_experiments folder. Running any of these files runs the corresponding experiment (which may take up to a few hours) and saves the results. The figures can then be generated using plot_experiment1.m, plot_experiment2.m, and plot_experiment3.m, respectively. The CSDL algorithm itself is implemented in learn_constrained_sparse_dictionary.m, and depends on ProjectOntoL1Ball.m (authored by John Duchi), sptoeplitz.m (authored by Tobin Driscoll), and multiconv.m.