Rethinking Developmental Curricula for Contrastive Visual Learning

TMLR Paper6602 Authors

Abstract: While large machine learning models have achieved remarkable results, they still fall short of the efficiency and adaptability characteristic of human perception. Inspired by infant visual development, we explore developmental curriculum learning strategies for contrastive learning, systematically isolating their effects under controlled conditions. Within a virtual environment, we modulated four dynamic factors, namely image blur, lighting complexity, avatar movement speed, and image complexity, to simulate developmental progression. However, none of these conditions improved downstream classification performance compared with a stable train setting. We then repeated the experiments on the real-world SAYCam dataset using dynamic movement speed and image complexity separately and obtained consistent results. These findings suggest that performance gains attributed to developmental learning do not arise directly from commonly assumed perceptual factors, which challenges the assumption that developmental-like progression inherently benefits learning and highlights the need for more principled curriculum design mechanisms. Our results offer a new perspective on curriculum design for self-supervised learning.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Steffen_Schneider1

Submission Number: 6602