An Information Criterion for Controlled Disentanglement of Multimodal Data

The repository contains the code for the DisentangledSSL method presented in the paper: An Information Criterion for Controlled Disentanglement of Multimodal Data (ICLR 2025). DisentangledSSL is a novel self-supervised approach for learning disentangled representations, separating information shared across different modalities and modality-specific information.

An example in vision-language domain:

An example in the biological domain:

Set up the environment

conda create -n multimodal python=3.10.9
conda activate multimodal
bash env.sh

Simulation Study

The code is in the synthetic_task/ folder. The command for DisentangledSSL (both step 1 and step 2) and baselines can be found in scripts/.

bash scripts/run_step1.sh
bash scripts/run_step2.sh

Citation

If you find this work useful in your research, please cite:

@article{wang2024information,
  title={An Information Criterion for Controlled Disentanglement of Multimodal Data},
  author={Wang, Chenyu and Gupta, Sharut and Zhang, Xinyi and Tonekaboni, Sana and Jegelka, Stefanie and Jaakkola, Tommi and Uhler, Caroline},
  journal={arXiv preprint arXiv:2410.23996},
  year={2024}
}