A Convexity of the self-supervised loss function