Causal Representation Learning Made Identifiable by Grouping of Observational Variables

Authors: Hiroshi Morioka, Aapo Hyvarinen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on synthetic data as well as a realistic high-dimensional image and gene regulatory network datasets show that our framework can indeed extract latent causal variables and their causal structure, with better performance than the state-of-the-art baselines.
Researcher Affiliation Academia Hiroshi Morioka 1 Aapo Hyv arinen 2 1RIKEN Center for Advanced Intelligence Project, Kyoto, Japan 2Department of Computer Science, University of Helsinki, Helsinki, Finland.
Pseudocode Yes Any optimization method can be used to minimize the loss (see Supplementary Algorithm 1 for example)...
Open Source Code Yes the implementation of G-Ca RL is available at https://github.com/hmorioka/GCa RL
Open Datasets Yes We used synthetic single-cell gene expression data generated by SERGIO (Dibaeinia & Sinha, 2020)... We also evaluated G-Ca RL on a more realistic observational model, by using a high-dimensional image dataset (3DIdent; Zimmermann et al. (2021)).
Dataset Splits No The paper mentions the total number of data points (e.g., "n was 2^16") but does not provide specific train/validation/test dataset splits (e.g., percentages or exact counts for validation).
Hardware Specification Yes convergence of a three-layer model by G-Ca RL took about 3 hours (Intel Xeon 3.6 GHz 16 core CPUs, 384 GB Memory, NVIDIA Tesla A100 GPU)
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, or other libraries).
Experiment Setup Yes The number of layers was selected to be the same as that of the observation model (L), and the number of units in each layer was 2dm except for the output (dm)... A maxout unit was used as the activation function in the hidden layers... The nonlinear functions were then trained by back-propagation with a momentum term (SGD) so as to optimize the cross-entropy loss of LR...