GMML2021Project

Project on Skoltech course "Geometric Methods in Machine Learning" on topic "Disentangled Representation Learning as Nonlinear ICA".

Gradual change of one feature, responsible for line thickness. NonlinearICA model on MNIST.

Authors

• Daniil Cherniavskii (Nonlinear ICA part)
• Ivan Matvienko (InfoGAN part)

Requirements installation

The code was tested using Python 3.7.6; installing necessary dependancies is done by:

pip3 install -r requirements.txt

Model training

To train Nonlinear ICA model and investigate what features may be mean, please refer to the Nonlinear ICA notebook.

For InfoGAN training, please refer to InfoGAN folder

Pretrained models

Already pretrained models for MNIST and CelebA can be found here.

Citation

@inproceedings{hyvarinen2019nonlinear,
  title={Nonlinear ICA using auxiliary variables and generalized contrastive learning},
  author={Hyvarinen, Aapo and Sasaki, Hiroaki and Turner, Richard},
  booktitle={The 22nd International Conference on Artificial Intelligence and Statistics},
  pages={859--868},
  year={2019},
  organization={PMLR}
}

@inproceedings{chen2016infogan,
  title={Infogan: Interpretable representation learning by information maximizing generative adversarial nets},
  author={Chen, Xi and Duan, Yan and Houthooft, Rein and Schulman, John and Sutskever, Ilya and Abbeel, Pieter},
  booktitle={Proceedings of the 30th International Conference on Neural Information Processing Systems},
  pages={2180--2188},
  year={2016}
}