Bayesian Batch Active Learning as Sparse Subset Approximation

Robert Pinsler, Jonathan Gordon, Eric T. Nalisnick, José Miguel Hernández-Lobato

Abstract

Leveraging the wealth of unlabeled data produced in recent years provides great potential for improving supervised models. When the cost of acquiring labels is high, probabilistic active learning methods can be used to greedily select the most informative data points to be labeled. However, for many large-scale problems standard greedy procedures become computationally infeasible and suffer from negligible model change. In this paper, we introduce a novel Bayesian batch active learning approach that mitigates these issues. Our approach is motivated by approximating the complete data posterior of the model parameters. While naive batch construction methods result in correlated queries, our algorithm produces diverse batches that enable efficient active learning at scale. We derive interpretable closed-form solutions akin to existing active learning procedures for linear models, and generalize to arbitrary models using random projections. We demonstrate the benefits of our approach on several large-scale regression and classification tasks.

1. Introduction

Active Learning (AL) is a critical technique for optimizing the labeling process in supervised learning. By selecting the most informative data points from an unlabeled pool, AL reduces the amount of labeled data required to achieve high performance. In the context of deep learning, sequential selection (one point at a time) is often computationally prohibitive due to the need for frequent model retraining. Batch Active Learning addresses this by selecting groups of points. However, naive batching strategies often select redundant points that are similar to each other, leading to inefficient use of labeling resources. This paper proposes a principled Bayesian framework that frames batch selection as a sparse subset approximation problem.

2. Methodology

The proposed approach, Active Bayesian CoreSets with Frank-Wolfe optimization (ACS-FW), treats batch selection as finding a small subset of the unlabeled pool that summarizes the information content of the entire pool.

2.1 Sparse Subset Approximation

The goal is to approximate the "ideal" posterior distribution—the one that would be obtained if the entire unlabeled pool were labeled—using only a small batch of points. This is inspired by coreset theory, which seeks to summarize large datasets while preserving statistical properties.

2.2 Optimization via Frank-Wolfe

To efficiently solve the subset selection problem, the authors employ the Frank-Wolfe algorithm. This projection-free optimization method greedily adds points to the batch to minimize the discrepancy between the batch's influence and the full pool's influence in a functional Hilbert space.

2.3 Scalability with Random Projections

For complex models like deep neural networks, calculating exact influences is difficult. ACS-FW utilizes random feature projections to map high-dimensional likelihoods into a lower-dimensional space, ensuring the algorithm scales linearly with the pool size.

3. Experimental Results

The authors evaluated ACS-FW across various benchmarks, including UCI regression tasks, the large-scale Year Prediction MSD dataset, and image classification benchmarks (CIFAR-10, SVHN, Fashion MNIST).

3.1 Diversity and Manifold Coverage

A significant advantage of ACS-FW is its ability to select diverse batches. Visualizations using t-SNE projections demonstrate that ACS-FW spreads queries across the data manifold, whereas traditional methods like MaxEnt or BALD tend to cluster queries in redundant regions.

Comparison of query diversity using t-SNE projections for MaxEnt, BALD, and ACS-FW (Ours).

3.2 Iterative Selection Dynamics

The iterative nature of the Frank-Wolfe algorithm ensures that once a point is selected, the residual error changes, making similar points less attractive for subsequent slots in the batch. This dynamic is contrasted with BALD, which continues to select points in the same high-uncertainty region.

Temporal evolution of selected points ($t=1$ to $t=10$) for BALD vs. ACS-FW.

3.3 Performance and Efficiency

ACS-FW consistently outperformed random sampling and naive batching, particularly on large-scale datasets. In image classification, it achieved performance competitive with state-of-the-art non-probabilistic methods like K-Center. Furthermore, the computational overhead for batch construction was found to be negligible compared to model training time.

4. Conclusion

ACS-FW provides a scalable and principled Bayesian approach to batch active learning. By framing the problem as a sparse subset approximation, it naturally incorporates diversity into the selection process. Future work may address the approximation error introduced by weight binarization and explore the impact of different posterior approximation techniques.

Summary updated

Abstract

Leveraging the wealth of unlabeled data produced in recent years provides great potential for improving supervised models. When the cost of acquiring labels is high, probabilistic active learning methods can be used to greedily select the most informative data points to be labeled. However, for many large-scale problems standard greedy procedures become computationally infeasible and suffer from negligible model change. In this paper, we introduce a novel Bayesian batch active learning approach that mitigates these issues. Our approach is motivated by approximating the complete data posterior of the model parameters. While naive batch construction methods result in correlated queries, our algorithm produces diverse batches that enable efficient active learning at scale. We derive interpretable closed-form solutions akin to existing active learning procedures for linear models, and generalize to arbitrary models using random projections. We demonstrate the benefits of our approach on several large-scale regression and classification tasks.
BibTeX
@inproceedings{pinsler2019bayesian,
  author = {Robert Pinsler and Jonathan Gordon and Eric T. Nalisnick and José Miguel Hernández-Lobato},
  title = {{Bayesian Batch Active Learning as Sparse Subset Approximation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=sFSbpAmBsX},
}
Figure