Geometric and Information Compression of Representations in Deep Learning

The project investigates the relationship between information compression and geometric properties of learned representations in deep neural networks, with a focus on conditional mutual information and neural collapse. This dataset contains checkpoints from the runs performed for empirical evaluation of the comparison between mutual information compression and geometric compression in the deep learning models.

In particular there are 5 separate tasks, with varying random seed initialization and training parameters. Checkpoints from every tenth epoch are stored. In order to load them one should use the code from the corresponding repository.