Questions for the seminar Paper " 	Collapse in (non-)contrastive learning"
-----------------------------------------------------------------------------------------------
Please send your answers to: marrakch@cs.uni-freiburg.de by 14:00 on 14.12.2022

1. Explain the expected collapse in non-contrastive self-supervised learning methods. How does it differ from the partial dimensional collapse? (~ 2-3 sentences)

2. To measure the redundancy in the learned feature embeddings, the authors introduce the cumulative explained variance. What is the main advantage of this approach over monitoring the standard deviation? (~ 2 sentences)

3. To fight against partial dimensional mode collapse, the authors suggest to use a continual training procedure. Explain briefly how it works. (~ 2-3 sentences)