Maximally Separated Active Learning

ECCV 2024 · Beyond Euclidean Workshop

Tejaswi Kasarla¹, Abhishek Jha¹, Faye Tervoort¹, Rita Cucchiara², Pascal Mettes¹

¹VIS Lab, University of Amsterdam²University of Modena and Reggio Emilia

Accuracy versus labelling budget curves on MNIST and TinyImageNet — Figure 1 from the paper.

Abstract

Active learning aims to select the most informative unlabeled samples for annotation. We propose Maximally Separated Active Learning, which applies the principle of maximum class separation — where class prototypes are arranged as an equiangular tight frame in feature space — as an inductive bias for sample selection. By querying samples that best support maximally separated representations, our approach achieves strong performance with fewer labeled examples across classification benchmarks.

BibTeX

@inproceedings{kasarla2024maximally,
  title={Maximally Separated Active Learning},
  author={Kasarla, Tejaswi and Jha, Abhishek and Tervoort, Faye and Cucchiara, Rita and Mettes, Pascal},
  booktitle={European Conference on Computer Vision Workshop},
  year={2024}
}