Speaker: David Lipshutz
Flatiron Institute
New York, NY
Title: Neuroscience-inspired Online Learning Algorithms
Abstract: A challenge in neuroscience is to understand the brain's learning algorithms and their implementations in neural circuits. Most learning in the brain is unsupervised or self-supervised. Many linear versions of un/self-supervised learning tasks can be expressed as generalized eigenvalue problems (e.g., canonical correlation analysis). I will discuss a mathematical framework for deriving online algorithms for learning tasks, including a broad class of generalized eigenvalue problems, that can be implemented in so-called "biologically plausible" neural circuit models which resemble neural circuits in the brain.