Robust Learning of a Single Neuron via Sharpness

Explore a 32-minute lecture on the robust learning of a single neuron through sharpness, presented by Jelena Diakonikolas from the University of Wisconsin-Madison. Delve into the world of cyclic block coordinate methods in continuous optimization, examining their historical significance, practical applications, and theoretical challenges. Discover how these methods, despite their widespread use in statistical learning software, have traditionally lacked strong theoretical foundations. Learn about a novel perspective that provides a more nuanced understanding of cyclic methods' non-asymptotic convergence. Uncover groundbreaking cyclic methods that demonstrate improved scaling with the number of blocks, breaking long-standing computational barriers. Gain insights into specific problem types where cyclic methods can match or surpass the performance of randomized or full-vector-update approaches, even in worst-case scenarios.