PiCCL: A lightweight multiview contrastive learning framework for image classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Self-supervised learning (SSL) is crucial for leveraging unlabeled data.
• Existing contrastive learning frameworks often use complex structures and loss functions.
• There is a need for simpler, more efficient SSL methods.

Purpose of the Study:

• Introduce PiCCL (Primary Component Contrastive Learning), a novel self-supervised contrastive learning framework.
• Develop a computationally lightweight and generalizable SSL method.
• Demonstrate PiCCL's effectiveness across various datasets and learning scenarios.

Main Methods:

• Utilized a multiplex Siamese network with multiple identical branches.
• Employed a simple image augmentation strategy to generate multiple positive samples.
• Designed a custom, computationally lightweight loss function (PiCLoss).

Main Results:

• Achieved top performance on CIFAR-10 (94%), CIFAR-100 (72%), and STL-10 (97%) datasets.
• Demonstrated superior performance in small batch learning scenarios (93% accuracy on STL-10 with batch size 8).
• Outperformed state-of-the-art self-supervised algorithms in benchmark tests.

Conclusions:

• PiCCL offers a simple, lightweight, and effective approach to self-supervised contrastive learning.
• The multiplex Siamese structure and custom loss function enhance learning efficiency and performance.
• PiCCL shows particular promise for resource-constrained environments and small batch learning.