A Simple Framework for Contrastive Learning of…

SimCLR turns two altered views of one photograph into a training target without class labels. Read it to see why easy visual shortcuts can spoil representation learning, why the projection head is discarded after training, and why the headline ImageNet result needs its model size attached.

Reading focus: How same-source image views define positives while other batch views supply competing negatives. Why augmentation composition and a nonlinear projection head change the representation that transfers. How to interpret linear evaluation, batch-size tradeoffs, and the wider model behind the headline accuracy.

ICML 2020. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 30 min read, easy difficulty.