MCDAL: Maximum Classifier Discrepancy for Active Learning.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• State-of-the-art active learning often uses Generative Adversarial Networks (GANs), which can be unstable and sensitive to hyperparameters.
• Existing methods for sample acquisition in active learning face challenges with stability and optimal data selection.

Purpose of the Study:

• To propose a novel and stable active learning framework, Maximum Classifier Discrepancy for Active Learning (MCDAL).
• To address the limitations of GAN-based methods in active learning by introducing a new sample acquisition strategy.

Main Methods:

• MCDAL utilizes prediction discrepancies between multiple classifiers to identify uncertain samples.
• Two auxiliary classification layers are employed to learn tighter decision boundaries by maximizing their prediction discrepancies.
• A novel acquisition function is developed based on these classifier discrepancies.

Main Results:

• The proposed MCDAL framework demonstrates superior performance compared to existing state-of-the-art methods.
• Empirical results show improved performance on image classification and semantic segmentation datasets in active learning scenarios.
• The method offers a stable and effective alternative to GAN-based active learning approaches.

Conclusions:

• MCDAL provides a robust and effective active learning framework by leveraging classifier prediction discrepancies.
• The approach offers a promising direction for improving sample selection in machine learning, particularly in computer vision tasks.
• MCDAL outperforms GAN-based methods and other contemporary techniques in active learning experiments.