Contrastive learning

A training method that teaches a model what belongs together by showing it matched and mismatched pairs, then pulling the matches closer together inside the model's representation and pushing the mismatches apart. CLIP, the model behind much of today's image and text AI, was trained this way on images paired with their captions. In healthcare the same trick works on chest X-rays paired with their reports: nobody has to label each finding by hand, because the pairing is the label. This is how most medical vision-language models get built without a hand-labeled dataset, and it is also where their blind spots come from, since a model only learns the distinctions its pairs happened to contain.

Every hospital already holds millions of matched pairs it never thought of as training data: an image and the report written about it. Contrastive learning turns that pairing into supervision, which is why chest radiography moved fastest, and it is also why such a model inherits the reporting habits of the archive it learned from. A finding your department words vaguely is a finding the model learns vaguely.

Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.