The Causal-Neural Connection: Expressiveness, Learnability, and Inference

Authors: Kevin Xia, Kai-Zhan Lee, Yoshua Bengio, Elias Bareinboim

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Simulations corroborate the proposed approach. In Sec. 5, we perform experiments with one possible implementation which support the feasibility of the proposed approach.
Researcher Affiliation Collaboration Kevin Xia Causal AI Lab Columbia University kmx2000@columbia.edu Kai-Zhan Lee Bloomberg L.P. Columbia University kl2792@columbia.edu Yoshua Bengio MILA Université de Montréal yoshua.bengio@mila.quebec Elias Bareinboim Causal AI Lab Columbia University eb@cs.columbia.edu
Pseudocode Yes Algorithm 1: Identifying/estimating queries with NCMs. Algorithm 2: Training Model
Open Source Code No The paper does not contain an explicit statement or a direct link to the source code for the methodology described in the paper. It only references a third-party library 'pytorch-made' [36].
Open Datasets No Observational data is generated from 8 different SCMs. The paper describes generating data but does not specify a publicly available or open dataset with concrete access information (link, DOI, formal citation).
Dataset Splits No The paper mentions training epochs and evaluating performance but does not provide specific details on dataset splits (e.g., percentages, sample counts for training, validation, or testing sets).
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper refers to general categories like "gradient descent tools" and mentions "pytorch-made" but does not provide specific version numbers for these or other software dependencies necessary for replication.
Experiment Setup Yes The parameter λ is set to 1 at the beginning, and decreases logarithmically over each epoch until it reaches 0.001 at the end of training. The classification accuracies per training epoch are shown in Fig. 4 (middle row) over 3000 training epochs. We rely on a hypothesis testing step such as |f(c M(θmax)) f(c M(θmin))| < τ for quantity of interest f and a certain threshold τ, with τ = 0.01 (blue), 0.03 (green), 0.05 (red).