ANC Workshop - Michael Gutmann and Benjamin Rhodes

Speaker: Michael Gutmann

Title: Neural Approximate Sufficient Statistics

Abstract:

My talk focuses on the core methodology of the ICLR paper https://openreview.net/pdf?id=SRDuJssQud

Specifically, I'll review the concept of sufficient statistics and explain that we can learn them by learning mutual information maximizing representations. I'll then explain how we used the learned statistics to boost the performance of both classical and recent methods for Bayesian parameter inference when the likelihood is intractable but sampling from the model is possible.

Speaker: Ben Rhodes

Title: Predicting bacterial virulence from whole genome sequences

Abstract: In this talk, we discuss the fitting of classifiers on entire bacterial genomes to predict virulence, and how such classifiers may be used to identify ‘significant’ regions of DNA that can then be experimentally investigated. We will focus on the two major issues: 1) how to compactly represent bacterial genomes, which are often millions of nucleotides in length and 2) what type of classifiers are appropriate, especially when the input dimensionality greatly exceeds the number of samples, and strong correlations are present between the features.