Understanding Label Bias in Single Positive Multi-Label Learning

Julio Arroyo, Pietro Perona, Elijah Cole

Keywords: computer vision, deep learning, single positive label, multi-label, label bias

TL;DR: Label bias matters for single positive multi-label learning.

Abstract: Annotating data for multi-label classification is prohibitively expensive because every category of interest must be confirmed to be either present or absent. Recent work on single positive multi-label (SPML) learning has shown that it is possible to train effective multi-label classifiers using only one positive label per image. The standard benchmarks for SPML are derived from traditional multi-label classification datasets by retaining one positive label for each training example (chosen uniformly at random) and discarding all other labels. However, in realistic annotation settings it is not likely that positive labels are chosen uniformly at random. In this work, we explore the effect of label bias in SPML.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2305.15584/code)