EasyDeepLearn

DINO — self-distillation with no labels.

hard

Answer

  • Student network learns to match teacher (EMA of student) on different augmentations of same image.
  • Uses ViT backbone → produces amazing self-attention maps (unsupervised object segmentation emerges!).
  • Centering + sharpening on teacher output prevents mode collapse.
  • DINOv2 (Meta 2023) scaled it to 1B parameters + 142M images → foundation vision encoder rivaling CLIP.
Check yourself — multiple choice
  • Random
  • Student matches EMA-teacher on augmented views; centering + sharpening prevents collapse; ViT backbone → emergent segmentation attention maps; DINOv2 at 1B params
  • Same as SimCLR
  • Not real

DINO: self-distillation from EMA teacher; emergent segmentation on ViT.

#representation-learning

Practise Unsupervised Learning

214 interview questions in this topic.

Related questions