Sinkhorn Divergence for Low-Budget Active Learning

We consider low-budget active learning, which consists of selecting a limited number of points, the coreset, such that a model can be trained to high accuracy on the selection only. This problem is particularly relevant in contexts where labeling requires costly expert intervention, as in medical applications. In this work, we propose using the Sinkhorn divergence as the coreset selection criterion in the optimization problem modeling active learning. The interest in the Sinkhorn divergence is twofold: it allows us to get dimension-free sample complexity results, it admits computationally efficient gradient evaluations, opening the way to using gradient-based algorithms with guarantees on the solution quality. Experiments on standard image and EEG benchmarks show that our method outperforms state-of-the-art heuristics in low-budget settings.