Research map / Label-efficient and robust learning / Recognition and long-tail

Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Deep active learning · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Composite Active Learning (CAL), a multi-domain active learning method that uses learned similarity weights and a conditional discriminator to allocate labeling budget across domains, with a theoretical error bound.

Problem

Active learning under a fixed labeling budget has been studied only in the single-domain setting; many real tasks involve multiple domains, and it is unclear how to allocate labeling budget across them.

Approach

Uses similarity weights between domains, learned end-to-end with an encoder, classifier and conditional discriminator in a min-max procedure, to guide the optimal labeling budget per domain and select informative instances; the paper claims a better error bound than current active learning methods.

Key results

• Authors report that CAL significantly outperforms state-of-the-art active learning methods across multiple datasets (specific numbers not given in the provided text).
• Authors report that CAL achieves a better error bound than current active learning methods, per their theoretical analysis.

Limitations

• Training time is higher than single-domain active learning because of the conditional discriminator
• Code not yet released at the time of rebuttal

Experimental setup

Baselines: single-domain active learning methods current active learning methods

Builds on: adversarial domain adaptation single-domain active learning

Closest papers by meaning

• Active Learning for Multiple Target Models (NeurIPS 2022 · similarity 0.89)
• A Cross-Domain Benchmark for Active Learning (NeurIPS 2024 · similarity 0.89)
• Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.87)
• Active Domain Adaptation with False Negative Prediction for Object Detection (CVPR 2024 · similarity 0.86)
• Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · similarity 0.86)
• One-shot Active Learning Based on Lewis Weight Sampling for Multiple Deep Models (ICLR 2024 · similarity 0.86)