On Mutual Information Maximization and Representation Learning
This is a companion Colab for the paper:
"On Mutual Information Maximization for Representation Learning"
Michael Tschannen*, Josip Djolonga*, Paul Rubenstein, Sylvain Gelly, and Mario Lucic
Click here to launch the Colab.
By default, the precomputed results will be loaded, but individual experiments can be run with the Colab by checking the
RUN_EXPERIMENTS checkbox. The batch size used in the paper was 128 and we average over 20 runs. With one run, the entire set of experiments will complete in ~2 hours. For multiple runs we suggest copying the code and running a stand-alone version. If you wish to run the experiments within the Colab, make sure you execute all cells in the "Setup" section.