On the Selection of Loss Functions Under Known Weak Label Models

Daniel Bacaicoa-Barber, Miquel Perelló-Nieto, Raúl Santos-Rodríguez, Jesús Cid-Sueiro

Abstract: This paper considers the problem of constructing proper loss functions for learning from weak labels by means of linear transformations of proper losses based on true labels. Recent works have shown that linear transformations defined by a left inverse of the transition matrix of the weak labelling process, transforms a true-label proper loss into a weak-label proper loss. In this paper, we show that the choice of both the true-label loss and the left inverse has a major influence on the performance of the learning algorithm, and we propose a novel method to optimize the loss selection. Some simulation results demonstrate the advantages of the proposed method.

External IDs:dblp:conf/icann/Bacaicoa-Barber21