Counterfactual Identifiability of Bijective Causal Models
Authors: Arash Nasr-Esfahany, Mohammad Alizadeh, Devavrat Shah
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our techniques in a visual task and demonstrate its application in a real-world video streaming simulation task. and We validate our techniques using a visual task ( 8.1) and demonstrate their application to a real-world system simulation task ( 8.2). and All results are in Table 2 with mean and standard deviation calculated over ten random seeds trained until training loss convergence. |
| Researcher Affiliation | Academia | Department of Electrical Engineering and Computer Science (EECS), Massachusetts Institute of Technology (MIT), Cambridge, MA 02139. Correspondence to: Arash Nasr-Esfahany <arashne@mit.edu>. |
| Pseudocode | No | The paper describes methods and processes but does not include any explicitly labeled pseudocode blocks or algorithm listings. |
| Open Source Code | Yes | Our code is available in github.com/counterfactual-BGM/cf2. |
| Open Datasets | No | For the ellipse generation task, the paper states 'We sample z PZ, u PU|Z=z and x PX|Z=z, and output the data tuple (z, x, v := f(u, x)). Here, PZ, PU|Z, and PX|Z are three predefined distributions (see Appendix E.1 for details).' For the video streaming simulation, it mentions 'We use the simulator to generate traces for each setting, which we use to learn the BGM.' and references Alomar et al. (2023) for the simulator, but does not provide a direct link or citation for a publicly available dataset used for training its own models. |
| Dataset Splits | No | The paper mentions training until 'training loss convergence' and using 'ten random seeds' for statistical analysis. However, it does not explicitly specify dataset splits (e.g., percentages or sample counts) for training, validation, or testing. |
| Hardware Specification | Yes | All experiments using A100 GPUs. |
| Software Dependencies | No | The paper mentions using 'Pyro (Bingham et al., 2018)' and 'Pytorch (Paszke et al., 2019)' but does not specify their version numbers. It also refers to 'linear rational splines (Dolatabadi et al., 2020)' and 'Adam (Kingma & Ba, 2015)' but not with explicit software versions. |
| Experiment Setup | Yes | We build all CGMs using NFs with linear rational splines... All splines we use have 16 bins for mapping ( 3, +3) to ( 3, +3). We use affine transforms at input and output layers to calibrate the range. Condition networks are all MLPs with two hidden layers, each with 64 units. We use batch size of 220 and run all experiments using A100 GPUs. We train all models using the default implementation of Adam (Kingma & Ba, 2015) in Pytorch (Paszke et al., 2019). |