MCMC Training of Bayesian Neural Networks

Explore Markov chain Monte Carlo (MCMC) methods for training Bayesian neural networks in this comprehensive lecture by Radford Neal from the University of Toronto. Delve into the background, outline, and key concepts of Bayesian neural networks, including nonbasing training, prior distributions, and symmetric stable distributions. Examine the benefits of Bayesian inference and the application of Hamiltonian Monte Carlo. Learn about the Flexible Bayesian Modeling Software and its practical applications in virus bioresponse prediction and CFAR 10 training. Gain insights into training validation errors and predictive performance, concluding with a Q&A session to address audience inquiries.

Syllabus

Introduction
Background
Outline
Bayesian Neural Networks
Nonbasing training
Bayesian approach
Prior distribution
Smooth functions
Symmetric stable distributions
Standard deviation
Hyperparameters
Prediction
Benefits
Bayesian inference
Markov chain Monte Carlo
Hamiltonian Monte Carlo
Flexible Bayesian Modeling Software
Virus Bioresponse
Training Validation Errors
Predictive Performance
CFAR 10 Training
Questions