MutexMatch: Semi-Supervised Learning With Mutex-Based Consistency Regularization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Semi-supervised learning (SSL) primarily focuses on leveraging unlabeled data.
• Existing SSL methods often overemphasize high-confidence samples, neglecting low-confidence ones.
• This underutilization of low-confidence data presents a key challenge in SSL.

Purpose of the Study:

• To introduce a novel approach, MutexMatch, for effectively utilizing low-confidence samples in SSL.
• To mitigate pseudo-labeling errors inherent in traditional SSL techniques.
• To enhance the overall performance of SSL models by fully exploiting unlabeled data.

Main Methods:

• Proposed MutexMatch, a mutex-based consistency regularization technique for SSL.
• Employed a true-positive classifier (TPC) for high-confidence samples to predict 'what it is'.
• Utilized a true-negative classifier (TNC) for low-confidence samples to predict 'what it is not', ensuring consistency of dissimilarity.

Main Results:

• MutexMatch demonstrated superior performance across various benchmark datasets including CIFAR-10, CIFAR-100, SVHN, STL-10, and mini-ImageNet.
• Achieved remarkable accuracy (92.23%) on CIFAR-10 with extremely limited labeled data (only 20 samples).
• Effectively addressed the challenge of pseudo-labeling errors by incorporating low-confidence samples.

Conclusions:

• MutexMatch offers a significant advancement in semi-supervised learning by novelly leveraging low-confidence samples.
• The method proves particularly effective in low-data regimes, showcasing its practical applicability.
• The consistency of dissimilarity approach enhances model robustness and performance in SSL tasks.