Reproduce “Discriminative Active Learning” on OpenResearch

About this paper

Daniel Gissin, Shai Shalev-Shwartz

Discriminative Active Learning (DAL) introduces a batch-mode active learning algorithm that re-frames sample selection as a binary classification task to make labeled and unlabeled data distributions indistinguishable. This method demonstrates competitive performance with state-of-the-art techniques for deep learning while being conceptually task-agnostic, facilitating efficient reduction of labeling costs.

Key insights

• Active learning can be effectively re-conceptualized as a binary classification problem to identify informative samples, offering an alternative to traditional uncertainty or margin-based approaches.
• Utilizing a learned data representation makes active learning algorithms conceptually task-agnostic, allowing for broader applicability across different deep learning tasks.
• Empirical re-evaluation indicates that many batch-mode active learning methods, including uncertainty sampling, consistently outperform random sampling, correcting previous literature claims.

Summary by alphaXiv.