Auto-Pairing Positives Through Implicit Relation Circulation for Discriminative Self-Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Representation Learning

Background:

• Contrastive learning is a popular self-supervised learning framework for representation learning.
• Traditional contrastive methods require carefully designed positive pairs, limiting their application in complex scenarios.
• Scenarios like dense prediction, multi-modal data, and vision-language tasks often lack easily obtainable positive pairs.

Purpose of the Study:

• To introduce an auto-pairing mechanism, Implicit Relation Circulation (IRC), for discriminative self-learning frameworks.
• To overcome the limitation of explicit positive pair requirements in contrastive learning.
• To enable contrastive learning in scenarios where positive pairs are difficult to obtain.

Main Methods:

• IRC employs a random walk across feature groups for a complex task (Task C) without explicit pairings.
• A simple task (Task S) with easily obtainable pairs is used to link the start and end of the random walk, forming a circulation.
• Cycle consistency is utilized as a supervision signal to automatically learn positive pairs within the random walk.

Main Results:

• IRC demonstrates effectiveness in learning dense image pixel relations, 3D point cloud relations, and image representations with language.
• Significant performance improvements were observed across various tasks, including semantic segmentation (3.0 mIoU), 3D segmentation (1.5 mIoU), 3D detection (1.3 mAP), and image classification (1.2 top1 accuracy).
• These gains were achieved with minimal parameter and computational overhead, showcasing IRC's efficiency.

Conclusions:

• Implicit Relation Circulation (IRC) is a versatile, plug-and-play mechanism that enhances discriminative self-learning.
• IRC successfully addresses the challenge of positive pair acquisition in contrastive learning.
• The proposed method offers a novel approach for representation learning, expanding its applicability to diverse and challenging domains.