Dr. Aditya Menon (Data61/Australian National University, Australia)
The hidden talents of logistic regression: from noisy labels to point processes
Class-probability estimation (CPE) techniques such as logistic regression are a key component in any multi-class classification toolkit. This talk aims to illustrate that these techniques can address a far broader spectrum of (un)supervised learning problems. We first show that CPE implicitly performs, in a precise sense, density ratio estimation. Inspired by the ubiquity of the latter, we then detail how CPE can address several distinct learning problems which have received interest of late, including coping with label noise in binary classification, learning classifiers with a fairness constraint, and estimating parameters of a point process.
|Date||November 7, 2017 (Tue) 17:00 - 18:00|