Partial Label Learning via Self-Paced Curriculum Strategy

Abstract

Partial-Label Learning (PLL) aims to learn from the training data, where each example is associated with a set of candidate labels, among which only one is correct. Existing PLL methods to deal with such problem usually treat each training example equally and few works take the complexities of training examples into consideration. In this paper, inspired by the human learning mode that gradually learns from “easy” to “hard”, we propose a novel S elf- P aced C urriculum strategy based P artial- L abel L earning ( SPC-PLL ) algorithm, where curriculum strategy can predetermine prior knowledge to adjust the learning priorities of training examples, while self-paced strategy can dynamically select “easy” training examples for model induction according to its current learning progress. The combination of such two strategies is analogous to “instructor-student-collaborative” learning mode, which not only utilizes prior knowledge flexibly but also effectively avoids the inconsistency between the predetermined curriculum and the dynamically learned models. Extensive experimental comparisons and comprehensive ablation study demonstrate the effectiveness of such strategy on solving PLL problem.