Smooth Normalizing Flows

Authors: Jonas Köhler, Andreas Krämer, Frank Noe

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

Reproducibility Variable Result LLM Response
Research Type Experimental The numerical experiments in this work are tailored to show the benefits of smooth flows over non-smooth flows. Therefore, we mainly compare mixtures of bump functions to neural spline flows [13], which exhibit state-of-the-art generative performance but whose densities are only first-order continuously differentiable. (Section 6)
Researcher Affiliation Academia Jonas Köhler Andreas Krämer Frank Noé Department of Mathematics and Computer Science, Freie Universität Berlin Department of Physics, Freie Universität Berlin Department of Chemistry, Rice University, Houston, TX {jonas.koehler, andreas.kraemer, frank.noe}@fu-berlin.de
Pseudocode No The paper describes the methods and formulations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper references existing software and libraries (e.g., 'OpenMM 7' [14], 'Pytorch' [40]) used in their work, but does not provide any explicit statement about releasing their own source code or a link to a repository for the methodology described.
Open Datasets No The paper mentions using a 'small molecule, alanine dipeptide' and 'test data from MD', which are common in the field. It also states that alanine dipeptide 'is described in the supplementary material'. However, it does not provide a direct link, DOI, or formal citation for accessing the specific dataset used in their experiments within the provided paper text.
Dataset Splits No The paper mentions 'holdout test set' and 'test data from MD' and refers to 'Test set metrics' in Table 1. It also states 'see supplementary material for details on the flow and training setup'. However, it does not provide specific percentages, sample counts, or clear definitions of the train/validation/test splits within the main paper text provided.
Hardware Specification No The paper mentions using 'GPUs' for speedup ('To obtain speedup on GPUs we suggest to generalize the classic binary search...'), but it does not specify any particular GPU models, CPU models, memory details, or other specific hardware configurations used for the experiments.
Software Dependencies Yes The paper references 'OpenMM 7' [14] in its bibliography and text, indicating a specific version number for this software.
Experiment Setup Yes As a proof-of-concept for the force matching loss, we trained smooth flows for alanine dipeptide using a 1000:1 weighting between the force matching residual and the negative log likelihood, see supplementary material for details on the flow and training setup. and The training of spline flows with inclusion of the force matching error (ωf > 0) was notoriously unstable... In contrast, all training runs with our smooth flows concluded successfully. and MD simulations... were equilibrated in each potential for 1 ps using a Langevin thermostat with 10/ps friction coefficient. and using the same 1 fs time step.