AutoAL: Automated Active Learning with Differentiable Query Strategy Search

TL;DR

AutoAL introduces a differentiable, bi-level framework for automated active learning by coupling FitNet, which models sample informativeness, with SearchNet, which learns to select AL strategies from a continuous relaxation of the candidate set. This probabilistic, gradient-based approach enables seamless integration of uncertainty-based and diversity-based AL methods, reducing computational cost while adapting to data distributions. The method demonstrates robust, superior performance across seven natural and medical image datasets, with extensive ablations confirming the importance of the loss-prediction module, SearchNet architecture, and the size of the candidate pool. By enabling automatic, data-driven strategy selection, AutoAL offers practical gains in labeling efficiency and generalizability across tasks and domains. The work provides a foundation for extending differentiable strategy search to broader AL scenarios and structured prediction settings.

Abstract

As deep learning continues to evolve, the need for data efficiency becomes increasingly important. Considering labeling large datasets is both time-consuming and expensive, active learning (AL) provides a promising solution to this challenge by iteratively selecting the most informative subsets of examples to train deep neural networks, thereby reducing the labeling cost. However, the effectiveness of different AL algorithms can vary significantly across data scenarios, and determining which AL algorithm best fits a given task remains a challenging problem. This work presents the first differentiable AL strategy search method, named AutoAL, which is designed on top of existing AL sampling strategies. AutoAL consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. For any given task, SearchNet and FitNet are iteratively co-optimized using the labeled data, learning how well a set of candidate AL algorithms perform on that task. With the optimal AL strategies identified, SearchNet selects a small subset from the unlabeled pool for querying their annotations, enabling efficient training of the task model. Experimental results demonstrate that AutoAL consistently achieves superior accuracy compared to all candidate AL algorithms and other selective AL approaches, showcasing its potential for adapting and integrating multiple existing AL methods across diverse tasks and domains. Code is available at: https://github.com/haizailache999/AutoAL.

Figures (5)

• Figure 1: Overall framework of differentiable query strategy search for automated active learning (AutoAL). AutoAL leverages labeled pool to train the FitNet and SearchNet in a bi-level optimization mode. Data samples with the largest search score are then selected from the unlabeled pool.
• Figure 2: Overall performance on seven benchmark datasets: natural image datasets (top) and medical image datasets (bottom).
• Figure 3: Ablation study on three components of AutoAL. 'ResNet Backbone': without the loss prediction module, updating only the ResNet. 'Loss Prediction Module': without updating the ResNet backbone, optimizing solely with the loss prediction module. 'ResNet+Loss Prediction': the full AutoAL pipeline.
• Figure 4: Ablation study on the size of the candidate pool in AutoAL strategy selection.
• Figure 5: AL strategy scores across different AL rounds on CIFAR-100 and OrganCMNIST datasets.