GPS: Guided Positive Sampling in Self-Supervised Learning
Abstract: This paper introduces Guided Positive Sampling Self-Supervised Learning (GPS-SSL), a method aimed at incorporating prior knowledge into Self-Supervised Learning (SSL) positive sample selection. Unlike current SSL methods relying solely on Data-Augmentations (DA) to generate positive samples, GPS-SSL creates a metric space aligning distances with semantic relationships and enabling informed positive sample selection through nearest neighbor sampling. A direct byproduct of GPS-SSL –and its core motivation– is the reduced importance of devising optimal DA recipes to learn performant representations. Since the proposed method solely alters the positive pair sampling, it can be coupled off-the-shelf with many SSL methods. Evaluation against baseline SSL methods on diverse datasets demonstrates the effectiveness of GPS-SSL, especially in scenarios with minimal DA; thus offering potential for further research on advancing SSL beyond careful DA design.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Bamdev_Mishra1
Submission Number: 3071
Loading