Self-Supervised Learning · 2021
Emerging Properties in Self-Supervised Vision Transformers
Caron et al. · DINO
Shows that self-supervised ViTs learn features that explicitly contain scene layout and object boundaries, without ever being told to.