Rectangular Flows for Manifold Learning

Authors: Anthony L. Caterini, Gabriel Loaiza-Ganem, Geoff Pleiss, John P. Cunningham

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now compare our methods against the two-step baseline of Brehmer and Cranmer [8], and also study the memory vs. variance trade-off. We use the real NVP [15] architecture for all flows, except we do not use batch normalization [24] as it causes issues with vjp computations.
Researcher Affiliation Collaboration Anthony L. Caterini University of Oxford & Layer 6 AI anthony@layer6.ai Gabriel Loaiza-Ganem Layer 6 AI gabriel@layer6.ai Geoff Pleiss Columbia University gmp2162@columbia.edu John P. Cunningham Columbia University jpc2181@columbia.edu
Pseudocode No We include the CG algorithm in Appendix B for completeness. The pseudocode itself is not within the provided text.
Open Source Code Yes Our code is available at https://github.com/layer6ai-labs/rectangular-flows.
Open Datasets Yes FMNIST) [57] than to MNIST [36] with a model trained on the former. We now turn our attention to the tabular datasets used by Papamakarios et al. [45] We also compare performance on the CIFAR-10 dataset [33]
Dataset Splits Yes We use early stopping with our FID-like score across all models. We also used likelihood annealing, with all experimental details again given in Appendix G.
Hardware Specification No The provided text does not contain specific details about the hardware used (e.g., GPU/CPU models, memory amounts).
Software Dependencies No The paper mentions "Py Torch [47] library" (Section 4.3) but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We include a detailed explanation of this phenomenon in Appendix E, along with all experimental details in Appendix G. We also used likelihood annealing, with all experimental details again given in Appendix G.