On the properties of variational approximations in statistical learning.
Title: On the properties of variational approximations in statistical learning.
Joint work with James Ridgway (Bristol) and Nicolas Chopin (ENSAE).
Speaker: Prof. Pierre Alquier . (ENSAE Paris Tech)
Date: Thursday, 29th October 2015
Time: 4pm – 5pm
Location: Room E0.01, Science East
Abstract:
PAC-Bayesian bounds were introduced by David McAllester (1998) in order to control the prevision risk associated with procedures of aggregation of estimators. Usually, these bounds lead to aggregation with exponential weights, with good optimality properties. These aggregates are usually computed through Monte Carlo methods. However, in many practical applications with large data, the computational cost of Monte Carlo methods is prohibitive, and it is tempting to replace these by a (faster) optimization algorithm. This is the idea of Variational Bayes techniques (VB). In this work, we prove, using PAC-Bayesian theorems, that VB methods are well founded, in the sense that the loss in prediction accuracy is negligible.
Series: Statistics
Social Media Links