Navigating, Restructuring and Reshaping Learned Latent Spaces

Watch a 52-minute conference talk exploring how modern machine learning architectures embed inputs into lower-dimensional latent spaces. Delve into research findings from Justin Solomon's group at the Centre International de Rencontres Mathématiques, focusing on developing lightweight algorithms for navigating, restructuring, and reshaping these learned latent spaces. Learn about practical machine learning applications including low-rank adaptation of large models, regularization for promoting local latent structure, and efficient training/evaluation of generative models. Recorded during the SIGMA thematic meeting on Signal, Image, Geometry, Modeling, and Approximation in Marseille, France, this presentation examines empirical results and emerging theories about how lower-dimensional codes capture data variation. Access additional features through CIRM's Audiovisual Mathematics Library including chapter markers, keywords, abstracts, bibliographies, and mathematical classifications for enhanced learning and reference.