Assignment 9: There… Are… Four… Labels!

Deadline: June 24th, 20:00

In this assignment, we will implement self-supervised learning models that learn features on unlabeled data. There is a starter notebook on E-Learning with extensive explanations, so we will not repeat them here. We just summarize the task briefly:

1. Train a classifier on a (CIFAR10) subset, which will likely not perform well/overfit significantly.
2. Train a self-supervised model of your choice. We offer some starter code on autoencoders or a rotation prediction task.
3. Use your self-supervised model as a basis to train another classifier on the small CIFAR subset. Since this model can use features derived from the much larger unlabeled set, we hope to achieve better performance.

6 CP Extra work

Also train the other self-supervised model that you did not pick above. You do not have to repeat the steps above, i.e. you do not need to also use this model to set up a classifier. We just want to you to have seen how to do both an autoencoder as well as a self-supervised classification task (like rotation prediction).