Skip to content

Latest commit

 

History

History
32 lines (21 loc) · 998 Bytes

SSL.md

File metadata and controls

32 lines (21 loc) · 998 Bytes

Self supervised Learning

  1. Jigsaw: Patches are shuffled, model predicts order of these patches, which are shuffled randomly.

  2. Tranferable to video.

  3. For language: Example is BERT.

  4. Difference between CV and NLP:

  • NLP , the signal is discrete.
  • CV, signal is high dimensional, continous.

Recent Successes

MoCo - Momentum Contrast for SSL. Train end to end, not finetune.

PIRL - Semantic Representations should be invariant under image transformations. Task - Jigsaw.

Contrastive Predictive Coding- Precit encodings below a certain image patch.

Sparse Overcomplete Representation of Image Data with a Multi Layer Conv Layer-

Idea: Use self supervision to learning sparse overcomplete representation of image.

  1. Put in image, get back same image.

Why sparsity ? An inductive bias only . small number of factors are resposible for a single data point. Why overcompleteness. Enables model flexibility. More robust to noise.

  • Sparse Coding and Dictionary Learning.