[Paper Review] Fast Rates in Pool-Based Batch Active Learning

Claudio Gentile, Zhilei Wang+1 more|arXiv (Cornell University)|Feb 11, 2022

Machine Learning and Algorithms4 citations

TL;DR

This paper proposes a stage-wise greedy algorithm for pool-based batch active learning that balances informativeness and diversity in label querying, achieving minimax optimal excess risk rates with only logarithmic retraining rounds. The method uses G-optimal design via greedy selection over exponentially shrinking margin regions, enabling fast convergence even under noisy data, with theoretical guarantees that match standard statistical rates while depending mildly on batch size.

ABSTRACT

We consider a batch active learning scenario where the learner adaptively issues batches of points to a labeling oracle. Sampling labels in batches is highly desirable in practice due to the smaller number of interactive rounds with the labeling oracle (often human beings). However, batch active learning typically pays the price of a reduced adaptivity, leading to suboptimal results. In this paper we propose a solution which requires a careful trade off between the informativeness of the queried points and their diversity. We theoretically investigate batch active learning in the practically relevant scenario where the unlabeled pool of data is available beforehand ({\em pool-based} active learning). We analyze a novel stage-wise greedy algorithm and show that, as a function of the label complexity, the excess risk of this algorithm matches the known minimax rates in standard statistical learning settings. Our results also exhibit a mild dependence on the batch size. These are the first theoretical results that employ careful trade offs between informativeness and diversity to rigorously quantify the statistical performance of batch active learning in the pool-based scenario.

Motivation & Objective

• To address the challenge of reduced adaptivity in batch active learning, where frequent oracle interactions are costly and redundant sampling degrades performance.
• To develop a theoretically grounded algorithm that maintains high statistical efficiency in pool-based active learning with minimal retraining rounds.
• To establish fast convergence rates (matching minimax optimal rates) for batch active learning under realistic, noisy data conditions.
• To extend theoretical guarantees to generalized linear and nonlinear function spaces, beyond traditional linear models.
• To design a method that automatically adapts to noise levels without prior knowledge, ensuring robust performance across varying data conditions.

Proposed method

• The algorithm uses a stage-wise greedy strategy to select batches of points by focusing on exponentially shrinking regions of the margin space, enhancing both informativeness and diversity.
• At each stage, it employs G-optimal design computed via a greedy approximation to maximize information gain while minimizing redundancy.
• Pseudo-labels are generated using sequences of linear classifiers trained on progressively refined subsets of the data, improving model generalization.
• The method adapts to noise levels by dynamically adjusting the number of stages, which remains logarithmic in the pool size.
• For constant batch size, the algorithm allows batch size to scale as $ T^\beta $ with $ \beta < 1 $, preserving optimal rates under noise-dependent scaling.
• Theoretical analysis extends to generalized linear models (e.g., logistic regression) and nonlinear function classes via VC-subgraph dimension bounds, ensuring generalization.

Experimental results

Research questions

• RQ1 Can a batch active learning algorithm achieve minimax optimal excess risk rates in a noisy, pool-based setting?
• RQ2 How can informativeness and diversity be balanced in batch selection to avoid redundant labeling while maintaining statistical efficiency?
• RQ3 What is the dependence of the convergence rate on the batch size and noise level in a practical, realizable learning setting?
• RQ4 Can theoretical guarantees be extended beyond linear models to generalized linear and nonlinear function spaces in batch active learning?
• RQ5 How many retraining rounds are necessary to achieve optimal performance, and can this be made adaptive to noise without prior knowledge?

Key findings

• The proposed algorithm achieves excess risk bounds that match the known minimax optimal rates in standard statistical learning settings, as a function of the total number of labels requested.
• The number of retraining stages is at most logarithmic in the pool size and automatically adapts to the noise level without requiring prior knowledge of it.
• Even with a constant batch size $ B $, the algorithm retains optimal convergence rates when $ B $ scales as $ T^\beta $ with $ \beta < 1 $, depending on noise.
• For generalized linear models (e.g., logistic regression), restricting attention to exponentially small margin regions leads to milder dependence on loss curvature, improving generalization.
• Theoretical analysis extends to nonlinear function classes via VC-subgraph dimension, providing the first such guarantees for batch active learning in the pool-based scenario.
• The method ensures a mild dependence on batch size, demonstrating that batch learning can be both efficient and statistically optimal in realistic, noisy settings.