#simclr

Question 1: Single choice

On a shape-classification task, NMA trains a linear classifier on top of different encoders (all with the same architecture) while keeping only a fraction of the labels. With 5% of the labels, the classifier on a SimCLR encoder (pretrained without labels on all images) stays above 90% accuracy, while an encoder trained jointly with the classifier on the labelled images drops to about 45%. What explains the gap?

SimCLR pretraining uses every image without labels, so the classifier only has to map useful features to classes1
SimCLR encoders have more parameters than supervised ones, so they need fewer labels2
The supervised encoder sees augmented images, which confuse it when labels are scarce3
A contrastive loss acts like label smoothing, which makes any classifier more accurate4

Timed test