BatchGFN: Generative Flow Networks for Batch Active Learning

TL;DR

BatchGFN introduces a generative flow network approach for efficient batch active learning, enabling near-optimal batch sampling with reduced computational complexity and potential scalability to real-world applications.

Contribution

It presents a novel generative flow network method for batch active learning that improves efficiency and effectiveness over traditional greedy approaches.

Findings

Enables sampling near-optimal batches with a single forward pass.

Reduces computational complexity compared to greedy algorithms.

Shows promising early results for scaling to real-world tasks.

Abstract

We introduce BatchGFN -- a novel approach for pool-based active learning that uses generative flow networks to sample sets of data points proportional to a batch reward. With an appropriate reward function to quantify the utility of acquiring a batch, such as the joint mutual information between the batch and the model parameters, BatchGFN is able to construct highly informative batches for active learning in a principled way. We show our approach enables sampling near-optimal utility batches at inference time with a single forward pass per point in the batch in toy regression problems. This alleviates the computational complexity of batch-aware algorithms and removes the need for greedy approximations to find maximizers for the batch reward. We also present early results for amortizing training across acquisition steps, which will enable scaling to real-world tasks.