Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Finding separatrices of dynamical flows with Deep Koopman Eigenfunctions
Authors: Kabir Dabholkar, Omri Barak
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our approach on synthetic benchmarks, ecological network models, and high-dimensional recurrent neural networks trained on either neuroscience-inspired tasks or fit to real neural data. |
| Researcher Affiliation | Academia | Kabir V. Dabholkar Faculty of Mathematics Technion Israel Institute of Technology Haifa, Israel 3200003 EMAIL Barak Rappaport Faculty of Medicine and Network Biology Research Laboratory Technion Israel Institute of Technology Haifa, Israel 3200003 EMAIL |
| Pseudocode | Yes | Algorithm 1 Train Koopman Eigenfunction Network |
| Open Source Code | Yes | Our code is available on Git Hub and we share an interactive description of the work and its extensions in a Uni Reps blog. Code attached in supplementary zip file. Code is also provided to train the RNNs used for reverse-engineering. |
| Open Datasets | Yes | 668D RNN fit to mouse neural activity: To demonstrate our method in a high-dimensional (see Appendix H for scaling results) and neuroscientifically relevant setting, we applied it to a recurrent neural network (RNN) trained to reproduce mouse neural activity from Finkelstein et al. [9].11D Ecological Dynamics: We next apply our method (Appendix G) to a high-dimensional ecological model: a generalized Lotka Volterra (g LV) system fit to genuslevel abundance data from a mouse model of antibiotic-induced Clostridioides difficile infection (CDI) [38]. |
| Dataset Splits | No | The paper uses models fit to existing data, for example, 'RNN trained to reproduce mouse neural activity from Finkelstein et al. [9]' and 'generalized Lotka Volterra (g LV) system fit to genuslevel abundance data from a mouse model of antibiotic-induced Clostridioides difficile infection (CDI) [38]'. The experimental setup focuses on sampling phase space for KEF training, not on train/test/validation splits of these underlying datasets for the KEF method itself. |
| Hardware Specification | Yes | We ran all experiments on a system with four Ge Force GTX 1080 GPUs with 10 Gbps of memory each. |
| Software Dependencies | No | We compute ฯ(x) using Pytorch s torch.autograd.grad... We use the Adam optimiser [41]... (No version numbers for Pytorch or specific Adam implementation version are provided). |
| Experiment Setup | Yes | A summary of all hyperparameters is provided in Table 1. Dynamical System N, Koopman eigenvalue ฮป, balance regularisation weight ฮณbal, batch-size B, training iterations T, learning rate ฮท, Res Net depth L and width dhid, number of Radial Basis Functions M. |