The neural networks trained in amortized variational inference for causal discovery (AVICI) infer causal structure from data based on a simulator of the domain of interest. By being trained on simulated data, the models can acquire realistic inductive biases from prior knowledge that is hard to cast as score functions or conditional independence tests used in classical causal inference.

This is a collective model card for the original and follow-up checkpoints of the paper (Lorch et al., 2022, NeurIPS 2022).

All models share the same architecture and training parameters and only differ in their synthetic training data distributions. The sampling distributions are specified in the config.yaml file in the respective model folder.

@article{lorch2022amortized,
  title={Amortized Inference for Causal Structure Learning},
  author={Lorch, Lars and Sussex, Scott and Rothfuss, Jonas and Krause, Andreas and Sch{\"o}lkopf, Bernhard},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  year={2022}
}