SCL: Self-supervised contrastive learning for few-shot image classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot learning (FSL) models struggle with generalization when trained on limited data.
• Developing robust FSL methods is crucial for real-world applications with scarce labeled samples.

Purpose of the Study:

• To propose a novel few-shot learning approach, Self-supervised Contrastive Learning (SCL).
• To enhance model representation and improve generalization capabilities in few-shot scenarios.

Main Methods:

• SCL integrates multiple self-supervision objectives: contrastive learning for sample discrimination and rotation prediction for diversity.
• A multitask learning environment assigns base and rotation class labels to training samples.
• Complex data augmentation strategies are employed, independent of base class information.

Main Results:

• The proposed SCL method simultaneously minimizes base class loss, contrastive distance loss, and rotation class loss.
• This simultaneous optimization enables learning of generic features crucial for improved novel class performance.
• SCL demonstrates superior performance compared to state-of-the-art methods on benchmark few-shot image classification datasets.

Conclusions:

• SCL effectively improves few-shot learning performance through integrated self-supervision.
• The combination of contrastive and rotation-based self-supervision enhances feature learning and generalization.
• The approach offers a promising direction for advancing few-shot image classification.