One-bit Active Query with Contrastive Pairs

Yuhang Zhang, Xiaopeng Zhang, Lingxi Xie, Jie Li, Robert C. Qiu, Hengtong Hu, Qi Tian

Introduction

Active learning (AL) is a powerful paradigm for building high-performing models by strategically selecting the most informative data points for human labeling. However, traditional AL faces a significant scaling bottleneck. As the number of categories ($C$) increases—such as the 1,000 classes in ImageNet—it becomes increasingly difficult and time-consuming for human annotators to identify the exact correct label, requiring $\log_2 C$ bits of information per sample.

To address this, researchers have explored "one-bit" supervision, which simplifies the task to binary "Yes/No" queries (e.g., "Is this a dog?"). While easier for humans, previous methods often failed to utilize "No" answers effectively. This paper introduces ObCP (One-bit Active Query with Contrastive Pairs), a framework that maximizes the utility of both positive and negative feedback through contrastive learning.

The ObCP Framework

ObCP integrates one-bit active learning with a contrastive learning objective to refine model decision boundaries efficiently. The process follows a cyclical four-stage approach:

  1. Uncertainty-Based Selection: The model identifies images where predictions have been "unstable" (high variance) over recent training epochs, suggesting they are near decision boundaries.
  2. One-Bit Query: For a selected image, the annotator is asked: "Does this image belong to the top-predicted class $c$?"
  3. Contrastive Optimization:
    • If "Yes": The image is added to the labeled pool and "pulled" toward other images of the same class in the feature space.
  4. If "No": The image is treated as a hard negative. A specialized Negative Loss ($L_{neg}$) "pushes" the image representation away from class $c$, teaching the model what the image is not.
  5. Joint Training: The model is updated using a combined loss function of supervised cross-entropy, modified contrastive loss ($L_{ctr}$), and the new negative loss.
The ObCP framework: One-bit queries are processed through Positive Pulling for "Yes" answers and Negative Pushing for "No" answers to refine the model.

Key Insights: The Value of "No"

In traditional AL, an incorrect prediction is simply a failure. In ObCP, a "No" provides a high-value "hard negative" sample. Because the model was confident enough to predict that class incorrectly, being told it is wrong provides a strong signal for refining decision boundaries. This binary nature of the query maps perfectly to the pull/push mechanics of contrastive loss, making one-bit supervision intrinsically suited for contrastive alignment.

Experimental Results

ObCP was evaluated on CIFAR-10, CIFAR-100, and ImageNet. The method consistently outperformed existing state-of-the-art AL methods like VAAL and Core-set across all benchmarks.

Mean accuracy on CIFAR-10: ObCP (orange star) significantly outperforms competing active learning methods as the percentage of labeled data increases.

On CIFAR-10, ObCP reached 92.6% accuracy using only 25% of labels, surpassing methods that required 40%. The results were even more striking on ImageNet, where the cost of traditional labeling is highest. With a budget of only 10% total information bits, ObCP achieved 61.3% accuracy, whereas the previous state-of-the-art (VAAL) required 30% of labels to reach only 53.3% accuracy.

Conclusion and Future Work

ObCP demonstrates that one-bit queries are not only easier for humans but, when paired with contrastive learning, are more information-efficient for models. While currently optimized for image classification, future work may extend this "Yes/No" paradigm to more complex spatial tasks like object detection and localized querying.

Summary updated

Abstract

How to achieve better results with fewer labeling costs remains a challenging task. In this paper, we present a new active learning framework, which for the first time incorporates contrastive learning into recently proposed one-bit supervision. Here one-bit supervision denotes a simple Yes or No query about the correctness of the model's prediction, and is more efficient than previous active learning methods requiring assigning accurate labels to the queried samples. We claim that such one-bit information is intrinsically in accordance with the goal of contrastive loss that pulls positive pairs together and pushes negative samples away. Towards this goal, we design an uncertainty metric to actively select samples for query. These samples are then fed into different branches according to the queried results. The Yes query is treated as positive pairs of the queried category for contrastive pulling, while the No query is treated as hard negative pairs for contrastive repelling. Additionally, we design a negative loss that penalizes the negative samples away from the incorrect predicted class, which can be treated as optimizing hard negatives for the corresponding category. Our method, termed as ObCP, produces a more powerful active learning framework, and experiments on several benchmarks demonstrate its superiority.
BibTeX
@inproceedings{zhang2022one,
  author = {Yuhang Zhang and Xiaopeng Zhang and Lingxi Xie and Jie Li and Robert C. Qiu and Hengtong Hu and Qi Tian},
  title = {{One-bit Active Query with Contrastive Pairs}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.1109/cvpr52688.2022.00947},
}
Figure