#representational-collapse

Question 1: Single choice

A SimCLR encoder is pretrained once with the usual many negative pairs per image and once with only 2. With few negatives, the features of most images become highly similar to each other, and a shape classifier on top can no longer tell ovals from squares. What role do the negative pairs play?

They push the features of different images apart; without enough of them, mapping everything alike satisfies the loss1
They provide labels for the classifier, so fewer negatives means fewer labelled examples2
They are extra augmentations, so fewer negatives means fewer augmented views of each image3
They only speed up training; after enough epochs two negatives give the same features4

#representational-collapse

Timed test