Please join us for a short reception after the seminar. 

Title: Fantastic Loss Functions and Where to Find Them

ABSTRACT: Throughout science and engineering, we use statistics and  machine learning to fit a model to data, where this "fit" is judged by some loss function.  But which loss function should one choose?  Every choice of loss function corresponds to a statistic of the data, the one it "elicits", and minimizing the loss over the data will pull the model toward this statistic.  For example, in ordinary least-squares regression, this statistic is the mean.  To understand how to design (fantastic) loss functions, therefore, we must understand which losses elicit which statistics.  In particular, given a desired statistic, we wish to design a loss that elicits it.  Unfortunately, some important statistics are not elicited by any loss function, and this impossibility is where our journey begins.  The talk will give an overview of loss  function design, introduce new techniques to design loss functions for non-elicitable statistics, and discuss ties to other areas of computer science and economics.

BIO: Rafael Frongillo is an Assistant Professor of Computer Science at CU Boulder.  His research lies at the interface between theoretical machine learning and economics, primarily focusing on domains such as information elicitation and crowdsourcing which involve the exchange of information for money, and drawing techniques from convex analysis, game theory,optimization, and statistics.  Before coming to Boulder, Rafael was a postdoc at the Center for Research on Computation and Society at Harvard University and at Microsoft Research New York, and received his PhD in Computer Science at UC Berkeley, advised by Christos Papadimitriou and supported by the NDSEG Fellowship.