Sparse Models

Presentation

The course introduces principles behind data transformation and optimization methods at the heart of automatic learning and data science, from the perspective of sparsity and robustness, applied to digital data compression (mp3, jpg) and representations by predictive models, making extensive use of algorithmic experimentation, intuition and history of science.

Learning outcomes

• To understand practical and theoretical motivations of optimization algorithms used in automatic learning and data science.
• To implement related algorithms in an adapted manner by understanding their meaning in relation to the problem at hand.
• To link different methods and implement them in a data processing flow.

Syllabus

The course presents a travel through data analysis and learning, via different sparse tools and methods, aimed at explaining observations by a reduced number of parameters: data metrics, descriptors and transformations (norms, vector bases and frames, wavelets); implementation through data compression algorithms (audio, image, video, text); extension to prediction models (statistical moments, linear and polynomial regressions, parsimonious or robust models).