Building One-Shot Semi-Supervised (BOSS) Learning Up to Fully Supervised Performance.

Area of Science:

• Computer Science
• Machine Learning
• Deep Learning

Background:

• Deep learning models typically require large labeled datasets for optimal performance.
• Labeling extensive datasets is time-consuming and costly.
• Semi-supervised learning offers a potential solution by leveraging unlabeled data.

Purpose of the Study:

• To introduce and validate a novel one-shot semi-supervised (BOSS) learning methodology.
• To demonstrate that BOSS learning can achieve performance comparable to fully supervised learning with minimal labeled data.
• To investigate the impact of class prototype refining, class balancing, and self-training on semi-supervised learning accuracy.

Main Methods:

• Development of a BOSS learning framework combining class prototype refining, class balancing, and self-training.
• Proposal of a technique for selecting iconic examples as class prototypes.
• Implementation of class balancing strategies to enhance semi-supervised learning outcomes.
• Evaluation using CIFAR-10 and SVHN datasets, analyzing per-class and total test accuracies.

Main Results:

• BOSS learning achieved 95% test accuracy on CIFAR-10 with one labeled sample per class, closely matching fully supervised performance (94.5%).
• On SVHN, BOSS learning reached 97.8% test accuracy, comparable to fully supervised learning (98.27%).
• Class balancing methods were shown to significantly improve accuracy, enabling self-training to reach fully supervised levels.

Conclusions:

• Deep neural network training does not necessitate labeling large datasets.
• The BOSS methodology offers an efficient alternative to fully supervised learning, especially when labeled data is scarce.
• The findings highlight the potential of one-shot semi-supervised learning for practical deep learning applications.