Contrastive Learning
A powerful self-supervised learning paradigm where a model is trained to actively pull similar data points closer together and push dissimilar ones apart.
Think of It Like This
Like organizing a messy photo album by simply tossing all pictures of dogs into one pile and aggressively throwing pictures of cats to the other side.
Instead of relying on human labels, contrastive models take an image and create augmented crops of it. The model learns that these two crops should share an identical embedding, while explicitly learning that they must be far away from a completely different random image. This technique forms the mathematical foundation of powerful embedding models like CLIP.