What did SimCLR find about its projection head?

Question: In SimCLR, why is a learnable nonlinear projection placed between the encoder representation and the contrastive objective? A) Optimising the contrastive loss in a separate projected space can improve the quality of the underlying representation used downstream B) The projection guarantees that every representation dimension has unit variance and zero covariance C) The projection converts contrastive learning into maximum-likelihood density estimation D) It prevents false negatives by mapping semantically identical samples to exactly the same vector Correct: A Explanation: SimCLR found that inserting a nonlinear transformation before applying the contrastive loss substantially improved learned representation quality. This does not create guarantees about covariance, density estimation, or false-negative removal. Topic: advanced ML / contrastive learning / SimCLR