Abstract
The performance of deep neural networks improves with more annotated data, but the budget for annotation is often limited. Active learning (AL) addresses this by having a model request labels for data it finds most uncertain. However, many existing AL methods are task-specific or computationally expensive. We propose a novel, task-agnostic active learning method that attaches a small "loss prediction module" to a target network. This module learns to predict the loss of unlabeled inputs, allowing it to suggest the most informative data points for annotation. Validated across image classification, object detection, and human pose estimation, our method consistently outperforms state-of-the-art baselines.
Introduction
Deep learning models are notoriously data-hungry, yet high-quality human annotation remains a significant bottleneck. Active Learning aims to optimize the labeling process by selecting the most informative samples from an unlabeled pool. Traditional AL strategies often rely on task-specific heuristics, such as entropy for classification, which do not easily generalize to regression tasks like pose estimation. Furthermore, methods requiring multiple forward passes or model ensembles are often too slow for large-scale applications. The authors sought a universal metric for informativeness that works regardless of the specific task.
Methods: The Loss Prediction Module
The core innovation is a lightweight, parametric module designed to estimate the potential loss of an unlabeled sample.
Architecture
The loss prediction module is a small multi-layer perceptron (MLP) that extracts mid-level feature maps from various stages of the target network (e.g., a ResNet). By aggregating features from different depths, the module captures both low-level textures and high-level semantic information to judge if an image is difficult.
Ranking Loss
Training the module to regress absolute loss values is problematic because the target model's loss naturally decreases as it learns. To overcome this "moving target" problem, the authors employ a Ranking Loss. Instead of predicting an exact value, the module is trained on pairs of samples to determine which one will result in a higher loss. This relative approach ensures stability throughout the training process.
Experimental Setup
The approach was validated on three distinct tasks to prove its versatility:
- Image Classification: CIFAR-10 (10 classes).
- Object Detection: PASCAL VOC 2007 + 2012 (hybrid task of classification and bounding box regression).
- Human Pose Estimation: MPII (a pure regression task involving localizing body joints).
Key Results
The proposed "Learn Loss" method consistently outperformed standard baselines, including random sampling, entropy-based sampling, and core-set (diversity-based) approaches.
- CIFAR-10: Reached 91.01% accuracy, surpassing entropy (90.59%) and core-set (90.10%).
- Object Detection: Achieved 73.38% mAP, a significant jump over the random baseline (71.17%).
- Pose Estimation: Attained 80.46% [email protected]. Notably, while entropy-based methods struggled with this regression task, the task-agnostic approach remained highly effective.
- Correlation: Visualization showed that the predicted loss had a correlation coefficient of 0.68 with the actual loss, significantly higher than the 0.45 correlation seen with entropy.
Conclusion and Future Work
By treating loss as a universal proxy for informativeness, this framework provides a simple yet effective tool for active learning. The module is computationally efficient and generalizes across diverse computer vision tasks. A current limitation is the lack of explicit diversity; the model might select many similar "difficult" images. Future research will explore combining this uncertainty-based selection with density-based methods to balance informativeness with representativeness.
Summary updated