gonlairo/ssl_identifiability

Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style

Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style [NeurIPS 2021]

Official code to reproduce the results and data presented in the paper Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style.

Numerical data

To train:

> python main_mlp.py --style-change-prob 0.75 --statistical-dependence --content-dependent-style

To evaluate:

> python main_mlp.py --style-change-prob 0.75 --statistical-dependence --content-dependent-style --evaluate

Causal3DIdent Dataset

You can access the dataset here. The training and test datasets consists of 250000 and 25000 samples, respectively.

High-dimensional images: Causal3DIdent

To train:

> python main_3dident.py --offline-dataset OFFLINE_DATASET --apply-random-crop --apply-color-distortion

To evaluate:

> python main_3dident.py --offline-dataset OFFLINE_DATASET --apply-random-crop --apply-color-distortion --evaluate

BibTeX

@inproceedings{vonkugelgen2021self,
  title={Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style},
  author={von Kügelgen, Julius and Sharma, Yash and Gresele, Luigi and Brendel, Wieland and Schölkopf, Bernhard and Besserve, Michel and Locatello, Francesco},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021}
}

Acknowledgements

This repository builds on the following codebase. If you find the dataset/code provided here to be useful, I would recommend you to also cite the following,

@article{zimmermann2021cl,
  author = {
    Zimmermann, Roland S. and
    Sharma, Yash and
    Schneider, Steffen and
    Bethge, Matthias and
    Brendel, Wieland
  },
  title = {
    Contrastive Learning Inverts
    the Data Generating Process
  },
  booktitle = {Proceedings of the 38th International Conference on Machine Learning,
    {ICML} 2021, 18-24 July 2021, Virtual Event},
  series = {Proceedings of Machine Learning Research},
  volume = {139},
  pages = {12979--12990},
  publisher = {{PMLR}},
  year = {2021},
  url = {http://proceedings.mlr.press/v139/zimmermann21a.html},
}