Flows for simultaneous manifold learning and density estimation
Authors: Johann Brehmer, Kyle Cranmer
NeurIPS 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In a range of experiments we demonstrate how M-flows learn the data manifold and allow for better inference than standard flows in the ambient data space. |
| Researcher Affiliation | Academia | Johann Brehmer and Kyle Cranmer New York University EMAIL, EMAIL |
| Pseudocode | No | No structured pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | Yes | The code used in our study is available at http://github.com/johannbrehmer/manifold-flow. |
| Open Datasets | Yes | We generate these with a Style GAN2 [25] model trained on the FFHQ dataset [26], sampling n of the GAN latent variables while keeping all others fixed. ... In addition, we use the real-world Celeb A-HQ dataset [26]. |
| Dataset Splits | No | The paper mentions training and test data, but does not explicitly provide details for a separate validation split or how it was used beyond general evaluation. |
| Hardware Specification | No | Funding disclosure: This work was supported in part through the NYU IT High Performance Computing resources, services, and staff expertise. (This is a general statement and lacks specific hardware details like GPU/CPU models.) |
| Software Dependencies | Yes | We are grateful to the authors and maintainers of DELPHES 3 [32], ... PYTHIA8 [39], ... PYTORCH [40], PYTORCH-FID [41], SCIKIT-LEARN [42], and SCIPY [43]. |
| Experiment Setup | Yes | All models are based on rational-quadratic neural spline flows [17]. For tabular datasets, we construct transformations f and h by alternating coupling layers with either random permutations or invertible linear transformations, using between 20 and 35 coupling layers depending on the dataset. For image data, f is based on a multi-scale architecture [5] with between 20 and 28 coupling layers across four levels interspersed with actnorm layers and 1 1 convolutions, closely following Refs. [17, 18]. |