A Sparsity Principle for Partially Observable Causal Representation Learning

Authors: Danru Xu, Dingling Yao, Sebastien Lachapelle, Perouz Taslakian, Julius Von Kügelgen, Francesco Locatello, Sara Magliacane

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we propose two methods that implement these theoretical results and validate their effectiveness with experiments on simulated data and image benchmarks, e.g. Causal3DIdent (von Kügelgen et al., 2021), that we modify to test our partial observability setting.
Researcher Affiliation Collaboration 1University of Amsterdam 2Institute of Science and Technology Austria 3Max Planck Institute for Intelligent Systems, Tübingen, Germany 4Samsung SAIT AI Lab, Montreal 5Mila, Université de Montréal 6Service Now Research 7Seminar for Statistics, ETH Zürich.
Pseudocode No The paper does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured algorithmic steps outside of narrative text.
Open Source Code Yes We provide all the code for our method and the experiments at https://github.com/danrux/sparsity-crl.
Open Datasets Yes Experiments on different simulated datasets and established benchmarks highlight the effectiveness of our approach in recovering the ground-truth latents. (...) e.g. Causal3DIdent (von Kügelgen et al., 2021) (...) We provide all the code for our method and the experiments at https://github.com/danrux/sparsity-crl.
Dataset Splits No The paper mentions 'generate images online until convergence' for the multiple balls dataset and samples from Causal3DIdent, but it does not specify explicit train/validation/test dataset splits with percentages, sample counts, or citations to predefined splits for reproducibility.
Hardware Specification No The paper mentions support in using 'the Dutch National Supercomputer Snellius' in the acknowledgments, but it does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for the experiments.
Software Dependencies No The paper mentions software and libraries like 'Cooper' and 'Py Game (Shinners, 2011)' but does not provide specific version numbers for these or other key software components, which is necessary for full reproducibility.
Experiment Setup Yes For the encoder g and the decoder ˆf, we use 7-layer MLPs with [10, 50, 50, 50, 50, 10] n units per layer, where n is the number of causal variables, with Leaky Re LU activations for each layer except the final one in the piecewise linear case. (...) In our experiments, we use ϵ = 0.01 or 0.001, details are provided in App. D.2. (...) We solve both problems with the extra-gradient version of Adam (Gidel et al., 2018). (...) Primal optimizer learning rate 1e-4, Dual optimizer learning rate 1e-4/2, Batch size 6144 (Table 4).