In Search of Invariance in Brains and Machines

Explore the limitations of deep convolutional neural networks and discover a novel approach to addressing their shortcomings in this thought-provoking lecture by Bruno Olshausen, Professor at U.C. Berkeley's Helen Wills Neuroscience Institute and School of Optometry, and Director of the Redwood Center for Theoretical Neuroscience. Delve into the challenges faced by current deep learning models, including their susceptibility to adversarial attacks and difficulty generalizing to real-world scenarios. Learn about an innovative mathematical framework based on Lie theory that aims to overcome these issues by modeling hierarchical structures and describing transformations in object perception. Gain insights into unsupervised learning of shapes and their transformations using Lie Group Sparse Coding, and explore the potential of the generalized bispectrum in creating complete and robust invariant representations. This hour-long presentation offers a fresh perspective on improving machine learning algorithms by drawing inspiration from neuroscience and advanced mathematical concepts.