Interactive lesson~15 minIntermediate
Contrastive Learning
Contrastive learning teaches representations by comparing examples. Similar views are pulled together; unrelated examples are pushed apart.
SimCLRMoCoSelf-supervised
Mental model
Learn by saying “these two are the same idea” and “those are not.”
Self-supervised visual, text, audio, and multimodal systems often begin with contrastive objectives.
Invariance
balanced70% modeled signal
Separation
balanced65% modeled signal
Transfer quality
balanced54% modeled signal
Concept pipeline
Build the idea in four moves
Interactive lab
Build a self-supervised representation learner.
Augment
Create two views of the same underlying item.
Focus lens
The part that usually clicks late
Positive pairs
Good augmentations preserve identity while changing surface details.
Invariance
70
Separation
65
Transfer quality
54
Knowledge check
What does contrastive learning optimize?
Next horizon
Where this topic is headed
CLIP-style multimodal training
BYOL without negatives
Hard-negative mining