Neural networks: optimization, transition to linearity and deviations therefrom
Misha Belkin
University of California, San Diego
Instructor: Misha Belkin
Associate Professor in the Department of Computer Science and and Engineering and Statistics, Ohio State University.
Fit without Fear: an Interpolation Perspective on Optimization and Generalization in Modern Machine Learning
A striking feature of modern supervised machine learning is its consistent use of techniques that interpolate the data. Deep networks, often containing several orders of magnitude more parameters than data points, are trained to obtain near zero error on the training set. Yet, at odds with most theory, they show …