Speaker:

Matt Parry, Department of Mathematics and Statistics

Title:

Score matching and local scoring rules: at the interface of machine learning and statistics

Location:

Owheo G34 - 1:00 pm, Friday 7 September

Abstract:

There are a number of probabilistic models in machine learning, statistics and signal processing for which the normalization constant is unknown or cannot be computed in any practical way. Score matching is a technique, first fully developed by HyvŠrinen in 2005, to estimate parameters in such cases. Shortly after it was proposed, it was realized that score matching could be understood as arising from a new class of scoring rules. Scoring rules themselves are principled ways to evaluate probabilistic forecasts and have been developed by statisticians since the 1950s. In this talk, I will discuss recent developments in score matching, including penalized approaches, and applications to circular distributions, graphical models, classification challenges and gene network discovery.