Interventional Causal Representation Learning

Authors: Kartik Ahuja, Divyat Mahajan, Yixin Wang, Yoshua Bengio

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically demonstrate the practical utility of our theory. From data generation mechanisms ranging from polynomials to image generation from rendering engine (Shinners et al., 2011), we show that interventional data helps identification. ... In this section, we analyze how the practical implementation of the theory holds up in different settings ranging from data generated from polynomial decoders to images generated from Py Game rendering engine (Shinners et al., 2011).
Researcher Affiliation Collaboration 1FAIR (Meta AI) 2Mila-Quebec AI Institute, Universit e de Montr eal 3University of Michigan 4CIFAR Senior Fellow and CIFAR AI Chair.
Pseudocode Yes Algorithm 1 Summarizing our two step approach for both the independence of support (IOS) and interventional data case.
Open Source Code Yes Also, the code repository can be accessed at: github.com/facebookresearch/Causal Rep ID.
Open Datasets No The paper describes its own data generation processes (e.g., 'Polynomial decoder data: The latents for the observational data are sampled from PZ' and 'Image data: For image-based experiments, we used the Py Game (Shinners, 2011) rendering engine'). It does not provide links or citations to pre-existing public datasets for direct access.
Dataset Splits Yes For experiments in Table 2, we only use observational data (D); while for experiments in Table 3, we use both observational and interventional data (D D(i)), with details regarding the train/val/test split described in Table 5. ... Table 5. Statistics for the synthetic poly-DGP experiments (Case Train Validation Test Observational (D) 10000 2500 20000 Interventional (D(I)) 10000 2500 20000) ... Table 10. Statistics for the synthetic image experiments (Case Train Validation Test Observational (D) 20000 5000 20000 Interventional (D(I)) 20000 5000 20000).
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments. It only describes the models and data generation.
Software Dependencies No The paper mentions software like 'Py Game (Shinners, 2011)' and 'scikit-learn (Pedregosa et al., 2011)' but does not specify version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes Model parameters and evaluation metrics. We follow the two step training procedures described in 7. ... For further details on data generation, models, hyperparamters, and supplementary experiments refer to the App. B. ... Hyperparameters. We use the Adam optimizer with hyperparameters defined below. We also use early stopping strategy, where we halt the training process if the validation loss does not improve over 10 epochs consecutively. Batch size: 16, Weight decay: 5 10 4, Total epochs: 200, Learning rate: optimal value chosen from grid: {10 3, 5 10 4, 10 4}.