Jacobian-based Causal Discovery with Nonlinear ICA
Authors: Patrik Reizinger, Yash Sharma, Matthias Bethge, Bernhard Schölkopf, Ferenc Huszár, Wieland Brendel
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that the proposed method can infer the causal graph on multiple synthetic data sets, and in most scenarios outperforms previous work. ... 5 Experiments ... We experimentally show that our proposed method can infer the DAG across multiple synthetic data sets. |
| Researcher Affiliation | Academia | Patrik Reizinger EMAIL University of Tübingen, Germany International Max Planck Research School for Intelligent Systems (IMPRS-IS) European Laboratory for Learning and Intelligent Systems (ELLIS) ... Yash Sharma EMAIL University of Tübingen, Germany International Max Planck Research School for Intelligent Systems (IMPRS-IS) ... Matthias Bethge EMAIL University of Tübingen, Germany ... Bernhard Schölkopf EMAIL Max Planck Institute for Intelligent Systems, Tübingen, Germany ... Ferenc Huszár EMAIL University of Cambridge, United Kingdom ... Wieland Brendel EMAIL Max Planck Institute for Intelligent Systems, Tübingen, Germany |
| Pseudocode | Yes | Algorithm 1 Algorithm for multivariable CD and determining the causal order π |
| Open Source Code | Yes | Corresponding author. Code available at: github.com/rpatrik96/nl-causal-representations |
| Open Datasets | Yes | We demonstrate that the proposed method can infer the causal graph on multiple synthetic data sets, and in most scenarios outperforms previous work. ... The recent work of Bagi et al. (2023) proposes a variational inference-based approach from interventional data, where the authors partition their latent space into invariant (content) and variant (style) features, which is a paradigm also found in CL, including identifiability guarantees and competitive performance on the Causal3DIdent dataset (Von Kügelgen et al., 2021). |
| Dataset Splits | No | The paper describes the generation of synthetic data for various scenarios and dimensions, e.g., "For the permuted SEM DGPs, we sample 6 different orderings and 5 seeds for each problem dimensionality {3; 5; 8; 10} the number of seeds is 10 for non-permuted and sparse SEMs." However, it does not specify explicit training/validation/test splits (e.g., percentages or counts) of this generated data for evaluation purposes. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types) used for running its experiments. |
| Software Dependencies | No | We calculate the Jacobian of the inference model with the autograd module of PyTorch (Paszke et al., 2019) in the forward pass and vectorize the operation with the recently released functorch library (Horace He, 2021). The specific version numbers for PyTorch and functorch are not provided. |
| Experiment Setup | Yes | Table 6: Hyperparameters for our experiments (5) Parameter Values bf 1 6-layer MLP Activation Leaky ReLU Batch size 6144 Learning rate 1e-4 Rd [0; 1]d Cp 1 mp 0 Cparam 0.05 mparam 1 p 1 τ (in LCL) 1 α 0.5 |