FractAL: Soft Strategy Selection for Batch-Mode Active Learning
ORIGINAL / Soft Strategy Selection for Batch-Mode Active Learning
To address the challenge of pre-selecting acquisition strategies in batch active learning, this work proposes soft strategy selection and FractAL, which uses influence-function-based attribution and online mirror descent to dynamically allocate batch budget across strategies without retraining. Across 7 benchmarks spanning classification, regression, and genetic perturbation effect prediction, FractAL matches or outperforms all baselines, offering a practical solution for few
01 ABSTRACT
The source notes that deploying active learning typically requires choosing an acquisition strategy before any labels, but strategy performance varies and existing selection methods spend many rounds exploring. The authors propose soft strategy selection—using multiple strategies within a single batch—and FractAL, which infers per-strategy reward via influence functions and computes budget shares with online mirror descent. Benchmarks show FractAL is no worse than random sampling and all baselines across 7 settings, concentrating budget on strong strategies.
02 KEY FINDINGS
- FractAL infers per-strategy reward using influence-function-based data attribution, requiring no retraining, then allocates batch budget via online mirror descent.
- Across 7 setups covering classification, regression, and genetic perturbation effect prediction, every existing strategy selection method underperforms random sampling on at least one setup, while FractAL matches or outperforms all baselines including random sampling on every setup.
- FractAL automatically concentrates budget on the strongest strategies and prunes the weakest, suitable for deployments where the optimal strategy is unknown in advance.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG