23 Contrastive Learning and SimCLR

Context Prediction

An early method for visual learning where the network must predict the spatial arrangement of two random image patches (e.g., which patch is “top-right”). To solve this, the network must “understand” object content

Contrastive Learning (SimCLR):

• Core Idea: Learn representation by comparing different views of the same image.
• The Mechanism:
  1. Take an image and apply two different augmentation (cropping, colour, jitter, blur, etc.)
  2. Pull Together: Use a loss function to maximise the agreement between these two views in latent space
  3. Push Apart: Minimise the agreement between views of different images in the same batch
• Kluger Hans Effect: A warning that models can “cheat” by basing decisions on spurious correlations (like chromatic aberration/colour shifts) rather than actual content. Countermeasures (like colour dropping) are often necessary.