Interaction Asymmetry: A General Principle for Learning Composable Abstractions

Official code for the paper Interaction Asymmetry: A General Principle for Learning Composable Abstractions.

Datasets

The Sprites, CLEVR6, and CLEVR-Tex datasets used in our experiments can be downloaded at the following link.

Experiments

The following commands can be used to run experiments using the regularized Transformer autoencoder from our paper.

Sprites:

python train_model.py --data "sprites" --encoder "transformer" --decoder "transformer" --alpha 0.05 --beta 0.05

CLEVR6

python train_model.py --data "clevr" --encoder "transformer" --decoder "transformer" --alpha 0.05 --beta 0.05

CLEVRTex

python train_model.py --data "clevrtex" --encoder "transformer" --decoder "transformer" --alpha 0.05 --beta 0.05

To use the baseline models in our experiments, the inputs for the encoder/decoder and hyperparameter values should be changed accordingly.

BibTeX

If you make use of this code in your own work, please cite our paper:

@article{brady2024interaction,
  title={Interaction Asymmetry: A General Principle for Learning Composable Abstractions},
  author={Brady, Jack and von K{\"u}gelgen, Julius and Lachapelle, S{\'e}bastien and Buchholz, Simon and Kipf, Thomas and Brendel, Wieland},
  journal={arXiv preprint arXiv:2411.07784},
  year={2024}
}