Reproduce “BatchGFN: Generative Flow Networks for Batch Active Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shreshth A. Malik, Salem Lahlou, Andrew Jesson, Moksh Jain, Nikolay Malkin, Tristan Deleu, Yoshua Bengio, Yarin Gal
BatchGFN introduces Generative Flow Networks to construct data batches for active learning, sampling points proportional to a batch-aware reward function such as Joint Mutual Information. This approach enables efficient batch acquisition with O(B) complexity at inference, achieving active learning performance on par with BatchBALD and showing promise for amortized training across acquisition steps.
Key insights
- Generative Flow Networks can be trained to sample complete batches of data points whose probability is proportional to a complex, non-additive batch reward function.
- This GFN-based approach enables efficient batch acquisition at inference time with O(B) complexity, unlike existing methods that incur higher costs.
- Amortization through "lookahead" training significantly reduces the retraining cost for the GFN as the underlying data distribution changes across active learning steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.