Reproduce “Active Domain Adaptation via Clustering Uncertainty-weighted Embeddings” on OpenResearch

About this paper

Viraj Prabhu, Arjun Chandrasekaran, Kate Saenko, Judy Hoffman

The paper introduces CLUE (Clustering Uncertainty-weighted Embeddings), a label acquisition strategy for Active Domain Adaptation that combines model uncertainty with feature-space diversity. This method consistently achieves higher target accuracy across various domain shifts and learning strategies, outperforming existing active learning and active domain adaptation approaches.

Key insights

• Pure uncertainty or diversity-based active learning methods are often ineffective for Active Domain Adaptation settings, highlighting the necessity of a combined approach.
• Using uncertainty-weighted clustering offers a principled way to balance informativeness and representativeness in label selection, effectively addressing the challenges posed by domain shifts.
• Integrating semi-supervised domain adaptation methods, such as Minimax Entropy (MME), with active label acquisition strategies significantly boosts performance on target domains.

Summary by alphaXiv.