A Fixed-Point Approach for Causal Generative Modeling

Authors: Meyer Scetbon, Joel Jennings, Agrin Hilmkil, Cheng Zhang, Chao Ma

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct an extensive evaluation of each method individually, and show that when combined, our model outperforms various baselines on generated out-of-distribution problems.
Researcher Affiliation Industry Work done when Joel Jennings and Cheng Zhang were affiliated with Microsoft Research. 1Microsoft Research 2Google Deep Mind. Correspondence to: Meyer Scetbon <tmscetbon@microsoft.com>.
Pseudocode Yes Algorithm 1 d-TOE(M, (Dtr, Gtr))
Open Source Code No The paper does not provide an explicit statement or link to its own open-source code for the described methodology. It only references implementations for baselines and other related works.
Open Datasets Yes We reproduce the procedure proposed in (Lorch et al., 2022) to generate synthetic datasets and their associated DAGs using randomly sampled SCMs.
Dataset Splits Yes Given a dataset D Rntot d where ntot is the total number of samples, we split it into three datasets w.r.t the sample size, with ratio 0.8, 0.1, 0.1 for training, validation and testing respectively.
Hardware Specification Yes We use 4 A100 GPUs with a total of 320 Gi B of memory and 85 CPUs to train our architecture M.
Software Dependencies No The paper mentions "the Adam implementation of Pytorch (Paszke et al., 2017)" but does not provide specific version numbers for Pytorch or any other critical software dependencies.
Experiment Setup Yes We consider TANM with an embedding dimension of D = 128, and L = 2 layers. The causal attention mechanism uses 8 heads with an embedding dimension of dhead = 32. We do not hyper-tune the model on each specific instance, but use the same training configuration for all experiments. See appendix E for details.