Loss function design for improved decision-making

Speaker: Jessie Finocchiaro (Harvard postdoc)

Talk Title: Loss function design for improved decision-making

Time & Location: 12/1, Friday, 1:30 - 2:30pm, SEC 1.413. 

Abstract: Algorithmic predictions are pervasive in our society, and these predictions are used to make decisions in settings ranging from banking to public health. This talk examines the relationship between the structure of downstream decision tasks and the design of algorithms: in particular, on the design of loss functions in supervised machine learning algorithms. Incorporating the decision structure into algorithm design helps us make "smarter errors" as necessary, but historical attempts to incorporate this structure are ad-hoc and labor-intensive. In this talk, I present a principled framework that, given a decision task, yields a convex loss function that lends itself to improved decision-making. This framework has been applied for top-k classification, image segmentation, max-margin prediction, and forming ordered partitions over outcomes.