Institutional tag:

Many useful empirical statistics, such as the sample variance and the Area Under the ROC Curve (AUC), are computed by averaging over all pairs of observations. Pairwise functions are also used as risk measures in machine learning problems like ranking, metric learning and graph inference. In this talk, I will present methods to efficiently estimate and optimize such functions in a decentralized network, where each agent holds a subset of the data. The proposed algorithms are asynchronous, rely on an efficient gossip protocol and enjoy convergence guarantees.

Dates:

Monday, January 30, 2017 - 14:00 to 15:15

Location:

Amphi Poirier, Bâtiment B, Centrale Lille

Speaker(s):

Aurélien Bellet

Affiliation(s):

CRIStAL et Inria Lille

Speaker's URL: