Explore the computational aspects of extremely wide neural networks in this 49-minute lecture by Yasaman Bahri from Google Brain. Delve into topics such as single-hidden layer neural networks of infinite width, Gaussian Process (GP) correspondences, and Bayesian inference with GP priors. Examine experimental results comparing neural network Gaussian process (NNGP) performance across hyperparameters, large depth behavior, and fixed points. Analyze phase diagrams, performance trends with width and dataset size, and empirical comparisons of various NN-GPs. Investigate the dynamics occurring in parameter space and gain insights into the best-performing networks, comparing GPs and SGD-trained neural networks. Part of the Frontiers of Deep Learning series at the Simons Institute.
Overview
Syllabus
Intro
Outline
Starting point
Single-hidden layer (shallow) neural networks of infinite width Consider a NN which
What GP does it correspond to?
Properties of the NNGP
Bayesian inference with a GP prior (Review)
Experiments from original work
Performance comparison
NNGP performance across hyperparameters
Large depth behavior & fixed points
Phase diagrams: experiments vs. theory
Performance trends with width and dataset size
Empirical comparison of various NN-GPS
Empirical trends
Best performing networks: comparison between GPs and SGD-NNS
Partway summary
What dynamics occurs in parameter space?
Closing Remarks
Taught by
Simons Institute
