What did SimCLR find about its projection head?

In SimCLR, why is a learnable nonlinear projection placed between the encoder representation and the contrastive objective?