Semi-Supervised Bayesian Active Learning with Task-Driven Representations

Kianoosh Ashouritaklimi, Tom Rainforth

Keywords: Active learning, Bayesian active learning, Representation learning

TL;DR: We propose a semi-supervised Bayesian active learning method that uses task-driven representations to overcome unsupervised representation failures on messy, uncurated data.

Abstract:

Current strategies for semi-supervised Bayesian active learning are generally based on learning unsupervised representations and then performing active learning on the resulting latent space with a supervised model. We find that this approach breaks down with messy, uncurated pools as the representations fail to capture the right similarities between our inputs. To address this, we propose the use of task-driven representations that are periodically updated during the active learning process. Our approach leads to more effective acquisitions and enhances model performance.

Submission Number: 33