Multi-Classifier Adversarial Optimization for Active Learning

Lin Geng, Ningzhong Liu, Jie QinAAAI 2023 · OpenReview

Abstract

Active learning (AL) aims to find a better trade-off between labeling costs and model performance by consciously selecting more informative samples to label. Recently, adversarial approaches have emerged as effective solutions. Most of them leverage generative adversarial networks to align feature distributions of labeled and unlabeled data, upon which discriminators are trained to better distinguish between them. However, these methods fail to consider the relationship between unlabeled samples and decision boundaries, and their training processes are often complex and unstable. To this end, this paper proposes a novel adversarial AL method, namely multi-classifier adversarial optimization for active learning (MAOAL). MAOAL employs task-specific decision boundaries for data alignment while selecting the most informative samples to label. To fulfill this, we introduce a novel classifier class confusion ($C^3$) metric, which represents the classifier discrepancy as the inter-class correlation of classifier outputs. Without any additional hyper-parameters, the $C^3$ metric further reduces the negative impacts of ambiguous samples in the process of distribution alignment and sample selection. More concretely, the network is trained adversarially by adding two auxiliary classifiers, reducing the distribution bias of labeled and unlabeled samples by minimizing the $C^3$ loss between classifiers, while learning tighter decision boundaries and highlighting hard samples by maximizing the $C^3$ loss. Finally, the unlabeled samples with the highest $C^3$ loss are selected to label. Extensive experiments demonstrate the superiority of our approach over state-of-the-art AL methods in terms of image classification and object detection.

Introduction

Deep learning's success is heavily dependent on large-scale annotated datasets. However, obtaining high-quality human labels is often prohibitively expensive and time-consuming. Active Learning (AL) addresses this by identifying the most informative samples for labeling, thereby maximizing performance with minimal annotation effort. While existing adversarial AL methods using Generative Adversarial Networks (GANs) or Variational Auto-encoders (VAEs) have shown promise, they often suffer from training instability and rely on simplistic discrepancy measures that ignore complex inter-class relationships.

The MAOAL Framework

To overcome these limitations, we propose Multi-Classifier Adversarial Optimization for Active Learning (MAOAL). The framework utilizes a feature generator ($G$), a primary classifier ($C$), and two auxiliary classifiers ($C_1$ and $C_2$).

Overview of the MAOAL framework, illustrating the interaction between the generator, classifiers, and the $C^3$ metric calculation for sample selection.

The core of MAOAL is the Classifier Class Confusion ($C^3$) metric. Unlike standard metrics that measure simple disagreement (like $l_1$ distance), $C^3$ computes a class correlation matrix from classifier outputs. This captures how classifiers confuse different categories, allowing the model to pinpoint truly ambiguous samples that sit in the confusing overlap between multiple categories near decision boundaries.

Training follows a "Min-Max" adversarial cycle:

  • Min Step: The generator ($G$) and primary classifier ($C$) minimize the $C^3$ loss to align the distributions of labeled and unlabeled data, ensuring the model generalizes well across the entire data pool.
  • Max Step: The auxiliary classifiers ($C_1$ and $C_2$) maximize the $C^3$ loss to push their decision boundaries apart, effectively creating a "sampling region" in the feature space that highlights the most informative, hard-to-classify samples.

Experimental Evaluation

We evaluated MAOAL on standard computer vision benchmarks including CIFAR-10, CIFAR-100, Caltech-101, and Pascal VOC for object detection.

Comparison of Mean Accuracy on the CIFAR-100 dataset. MAOAL (Ours) consistently outperforms other state-of-the-art active learning methods.

On the challenging CIFAR-100 dataset, MAOAL demonstrated a significant performance advantage over state-of-the-art (SOTA) methods like MCDAL and ADS. For instance, at a 25% labeling rate, MAOAL achieved a 1.97% higher accuracy than the nearest competitor. Similar trends were observed in object detection tasks on Pascal VOC, where MAOAL maintained a superior mean Average Precision (mAP) throughout the labeling cycle compared to both random sampling and other AL techniques.

Conclusion

MAOAL introduces a robust, hyper-parameter-free approach to active learning by leveraging inter-class correlations through the $C^3$ metric. By employing a triple-classifier adversarial strategy, it achieves more stable training and more effective sample selection than previous generative adversarial methods. While the current implementation shows strong results on standard backbones like ResNet-18, future work will explore scaling the framework to larger transformer-based architectures and diverse tasks such as semantic segmentation.

Summary updated

Abstract

Active learning (AL) aims to find a better trade-off between labeling costs and model performance by consciously selecting more informative samples to label. Recently, adversarial approaches have emerged as effective solutions. Most of them leverage generative adversarial networks to align feature distributions of labeled and unlabeled data, upon which discriminators are trained to better distinguish between them. However, these methods fail to consider the relationship between unlabeled samples and decision boundaries, and their training processes are often complex and unstable. To this end, this paper proposes a novel adversarial AL method, namely multi-classifier adversarial optimization for active learning (MAOAL). MAOAL employs task-specific decision boundaries for data alignment while selecting the most informative samples to label. To fulfill this, we introduce a novel classifier class confusion (C3) metric, which represents the classifier discrepancy as the inter-class correlation of classifier outputs. Without any additional hyper-parameters, the C3 metric further reduces the negative impacts of ambiguous samples in the process of distribution alignment and sample selection. More concretely, the network is trained adversarially by adding two auxiliary classifiers, reducing the distribution bias of labeled and unlabeled samples by minimizing the C3 loss between classifiers, while learning tighter decision boundaries and highlighting hard samples by maximizing the C3 loss. Finally, the unlabeled samples with the highest C3 loss are selected to label. Extensive experiments demonstrate the superiority of our approach over state-of-the-art AL methods in terms of image classification and object detection.
BibTeX
@inproceedings{geng2023multi,
  author = {Lin Geng and Ningzhong Liu and Jie Qin},
  title = {{Multi-Classifier Adversarial Optimization for Active Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=l4sw98e7aFC},
}
Figure