BOSS

BOSS implements one-shot semi-supervised learning to train image classifiers from a single labeled sample per class by leveraging unlabeled data to reach accuracies comparable to fully supervised models on CIFAR-10 and SVHN.

Key Features:

• One-Shot Semi-Supervised Learning: Trains models using only one labeled example per class together with unlabeled data to perform classification.
• Class Prototype Refining: Identifies and refines class prototypes via selection of iconic examples to improve class representation.
• Class Balancing: Applies class balancing techniques to mitigate imbalanced contributions from different classes during training.
• Self-Training: Uses iterative self-training where model predictions on unlabeled data are used as pseudo-labels to refine performance.
• Empirical Evaluation: Demonstrates test accuracies of 95% on CIFAR-10 and 97.8% on SVHN with one labeled sample per class, compared to fully supervised benchmarks of 94.5% and 98.27%, respectively.

Scientific Applications:

• Medical imaging: Enables training image classifiers when annotated medical images are scarce by leveraging unlabeled scans.
• Autonomous systems: Supports development of perception models for autonomous vehicles and robots with minimal labeled data.
• Low-label domains: Applies to any domain requiring efficient use of limited labeled resources for image classification.

Methodology:

Combines iconic example selection, class prototype refining, class balancing, and iterative self-training and is empirically evaluated on CIFAR-10 and SVHN.