Submitted on Jan 13, 2014
Author : XRCE
11. October, 2013
Speaker: Balázs Kégl, Researcher, CNRS, Université de Paris Sud, Orsay, France
Abstract: Constructing predictors in a sequential decision process is one of the possible ways to formalize budgeted learning. In the general setup, we train an agent which builds the prediction iteratively, making decisions on outputting the current prediction or gathering further information for improving it, based on rewards that mix predictive accuracy and cost of information. After sketching our motivating application of trigger design in experimental physics, we describe the MDDAG algorithm which is one of the possible instantiations of the sequential prediction paradigm. Making the predictor as lean as possible based on information gathered on a given instance leads to data-dependent feature selection similar to sparse coding. In the last part of the talk we explore this connection and argue that budgeted learning is not only a practical solution for real-time prediction, but may also be considered as a natural way to regularize predictors.