Speakers: Andrew Foong, David Burt, Javier Antoran
Abstract:
PAC -Bayes is a frequentist framework for obtaining generalisation error bounds. It has been used to derive learning algorithms, provide explanations for generalisation in deep learning, and form connections between Bayesian and frequentist inference. This reading group will cover a broad introduction to PAC bounds, the proof ideas in PAC -Bayes, and a discussion of some recent applications.
Suggested reading:
Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data: https://arxiv.org/abs/1703.11008
PAC -Bayesian Theory Meets Bayesian Inference: https://arxiv.org/abs/1605.08636
Learning under Model Misspecification: Applications to Variational and Ensemble Methods: https://arxiv.org/abs/1912.08335