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..
Characterizing control between interacting subsystems with deep Jacobian estimation
Authors: Adam J. Eisen, Mitchell Ostrow, Sarthak Chandra, Leo Kozachkov, Earl Miller, Ila Fiete
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that Jacobian ODE models outperform existing Jacobian estimation methods on challenging systems, including high-dimensional chaos. Applying our approach to a multi-area recurrent neural network (RNN) trained on a working memory selection task, we show that the sensory area gains greater control over the cognitive area over learning. |
| Researcher Affiliation | Collaboration | Adam J. Eisen Brain and Cognitive Sciences MIT Cambridge, MA 02139 EMAIL, Mitchell Ostrow Brain and Cognitive Sciences MIT Cambridge, MA 02139 EMAIL, Sarthak Chandra Brain and Cognitive Sciences MIT Cambridge, MA 02139 EMAIL, Leo Kozachkov IBM Thomas J. Watson Research Center IBM Research Yorktown Heights, NY 10598 EMAIL, Earl K. Miller Brain and Cognitive Sciences MIT Cambridge, MA 02139 EMAIL, Ila R. Fiete Brain and Cognitive Sciences MIT Cambridge, MA 02139 EMAIL |
| Pseudocode | No | The paper describes the methods in sections 3.1, 3.2, 3.3, and appendices A and D, but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | As part of the submission, we provide an anonymized link to download the analyzed data, as well as all code implementing the models, baselines, control analyses, and data generation. If the paper is accepted, the full code will be open sourced. |
| Open Datasets | Yes | All systems were simulated using the dysts package, which samples dynamical systems with respect to the characteristic timescale τ of their Fourier spectrum [99, 100]. ... As part of the submission, we provide an anonymized link to download the analyzed data, as well as all code implementing the models, baselines, control analyses, and data generation. |
| Dataset Splits | Yes | For all systems, the training data consisted of 26 trajectories of 12 periods, sampled at 100 time steps per τ. The validation data consisted of 6 trajectories of 12 periods sampled at 100 time steps per τ. The test data consisted of 8 trajectories of 12 periods sampled at 100 time steps per τ. ... For the training data, we generated 4096 random trials, and used 80% for training and the remainder for validation. |
| Hardware Specification | Yes | All models were able to be trained on a single H100 GPU, with 80 GB of memory. |
| Software Dependencies | No | All models were built in Py Torch [103]. Full implementation details are provided in appendices A and D. ... Path integration was performed using the trapezoid method from the torchquad package, with each integral discretized into 20 steps [101]. ODE integration was performed using the fourth-order Runge Kutta (RK4) method from the torchdiffeq package [102]. |
| Experiment Setup | Yes | All models were built in Py Torch [103]. Full implementation details are provided in appendices A and D. ... All models were implemented as four-layer MLPS, with the four layers having sizes of 256, 1024, 2048, and 2048 respectively. All models used a sigmoid linear unit activation. ... The batch size used was 16. Gradients were accumulated for 4 batches. Training epochs were limited to 500 shuffled batches. Validation epochs were limited to 100 randomly chosen batches. Testing used all testing data. Training was run for a maximum of 1000 epochs, 3 hours, or until the early stopping was activated (see Appendix D.8.6), whichever came first. |