Structured Neural Networks for Density Estimation and Causal Inference
Authors: Asic Chen, Ruian (Ian) Shi, Xiang Gao, Ricardo Baptista, Rahul G. Krishnan
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the efficacy of encoding structure into the learning process, we show that using Str NN its flow integrations to enforce a prescribed adjacency structure improves performance on density estimation and sample generation tasks. We experiment on both synthetic data generated from known structure equations and MNIST image data. |
| Researcher Affiliation | Academia | Asic Chen1 Ruian Shi1 Xiang Gao1 Ricardo Baptista2 Rahul G. Krishnan1 1University of Toronto, Vector Institute 2California Institute of Technology {asicchen, ruiashi, xgao, rahulgk}@cs.toronto.edu rsb@caltech.edu |
| Pseudocode | Yes | Algorithm 1: Greedy factorization |
| Open Source Code | Yes | The code to reproduce these experiments is available at https://github.com/rgklab/Structured NNs. |
| Open Datasets | Yes | We experiment on both synthetic data generated from known structure equations and MNIST image data. Details on the data generation process for all synthetic experiments can be found in Appendix C. [...] To study the effect of structure in image modeling, we use the binarized MNIST dataset considered in Germain et al. [2015], Salakhutdinov and Murray [2008]. |
| Dataset Splits | No | The paper uses standard datasets like MNIST and synthetic data, and mentions a 'test set' for evaluation, but it does not provide specific details on how the datasets were split into training, validation, and testing subsets (e.g., percentages, counts, or explicit splitting methodology). |
| Hardware Specification | No | The paper does not explicitly provide specific hardware details such as GPU or CPU models, memory specifications, or detailed computer configurations used for running the experiments. It only generally mentions 'Resources used in preparing this manuscript were provided in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute.' |
| Software Dependencies | No | The paper mentions software like the Gurobi optimizer (referencing 'Gurobi Optimization, LLC, 2023') and the Zuko software package (referencing 'Version 0.3.1' in its citation), but it does not explicitly list specific version numbers for *all* key software dependencies directly used in *their own* experiments (e.g., Python, PyTorch, or other specific libraries and their versions). |
| Experiment Setup | Yes | All discrete flow models use a UMNN [Wehenkel and Louppe, 2019] transformer and we grid-search other hyperparameters as described in Appendix E.2. [...] Interventions are performed by setting the intervened value α using one of eight integers perturbed from the mean of the intervened variable. [...] We generate 1000 observations using the synthetic SEM and derive counterfactual values x by posing queries with varying α values for each variable xj. |