Testing Noise Assumptions in Machine Learning Algorithms - Theory and Implementation

Watch a 35-minute lecture from the Joint IFML/MPG Symposium where Arsen Vasilyan from Simons Institute explores the fundamental question of efficiently testing whether training sets satisfy noise model assumptions in computational learning theory. Discover the first efficient algorithm for testing various noise assumptions on training data, extending the testable learning framework of Rubinfeld and Vasilyan. Learn about learning halfspaces over Gaussian marginals with Massart noise and understand the separation between classical learning with structured noise versus testable learning. Explore findings from joint research showing that testable learning for random classification noise requires super-polynomial time while classical learning remains trivial.