Do Labels Make AI Blind? Self-Supervision Solves the Age-Old Binding Problem

A recent paper demonstrates that Vision Transformers (ViTs) trained with self-supervised learning (SSL) methods naturally learn object binding, which is the ability to group visual elements into coherent objects. In contrast, models trained with traditional supervised classification using labels fail to develop this capability, instead learning shortcuts like recognizing textures. The research shows that SSL methods like DINO, MAE, and CLIP foster a more robust internal representation of visual scenes. This emergent property in self-supervised models provides insights into both artificial and biological vision, suggesting that relying on a single global loss from labels may be a fundamental weakness for creating robust AI systems.