Ishan Misra

Ishan Misra: Self-Supervised Deep Learning in Computer Vision | Lex Fridman Podcast #206

Ishan Misra, Lex Fridman

This discussion defines self-supervised learning as a paradigm that eliminates human annotation by extracting supervision signals directly from data structures, utilizing techniques like masking in natural language processing and contrastive learning in computer vision. Key speakers highlight how recent architectural advances in Transformers and non-contrastive methods enable scaling to billion-parameter models, though significant challenges remain regarding data efficiency, catastrophic forgetting, and the gap between pattern recognition and complex reasoning. The conversation concludes with applications in autonomous driving, philosophical implications for artificial consciousness, and practical guidance for researchers emphasizing hands-on implementation and focused problem selection.