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..
Modeling Neural Activity with Conditionally Linear Dynamical Systems
Authors: Victor Geadah, Amin Nejatbakhsh, David Lipshutz, Jonathan Pillow, Alex H Williams
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In example applications, we apply CLDS to model thalamic neurons that nonlinearly encode heading direction and to model motor cortical neurons during a cued reaching task. 3 Experiments Metrics For a given trajectory {y1:T , u1:T }, we denote as data reconstruction the mean emission E [y1:T | ˆx1:T , F(u1:T )] = (C(u1)ˆx1, . . . , C(u T )ˆx T ) from a the posterior mode ˆx1:T , computed with Kalman smoothing, given the observations y1:T and parameters F(u1:T ). As our primary metric, we use co-smoothing [28] to evaluate the ability of models to predict held-out single-neuron activity. Specifically, for the top 5 neurons with highest variance from the test set, we compute the coefficient of determination R2 between the true and reconstructed single-neuron firing rate, obtained by performing data reconstruction using only the other neurons. |
| Researcher Affiliation | Academia | 1Program in Applied and Computational Mathematics, Princeton University, Princeton NJ 2Center for Computational Neuroscience, Flatiron Institute, New York City NY 3Department of Neuroscience, Baylor College of Medicine, Houston TX 4Princeton Neuroscience Institute, Princeton NJ 5Center for Neural Science, New York University, New York City NY |
| Pseudocode | No | The paper describes the inference process using mathematical derivations and descriptive text, notably in Section 2.3 'Inference' and Appendix A.1, but it does not present a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | 1Our CLDS implementation is available at https://github.com/neurostatslab/clds. |
| Open Datasets | Yes | We analyze neural recordings of antero-dorsal thalamic nucleus (ADn) from Ref. [32] of the mouse HD system in mice foraging in an open environment (Fig. 3a). The data was accessed through the pynapple (PYthon Neural Analysis Package, https://pynapple.org/, MIT License) package, with the data itself stored in NWB format on OSF at https://osf.io/jb2gd. Finally, we analyzed neural recordings of dorsal premotor cortex (PMd) in macaques performing center-out reaching task (Fig. 4a) from Ref. [33] the associated article is under a CC BY 4.0 license. |
| Dataset Splits | Yes | The hyper-parameters {L, κ, σ} for the GP priors and the dimensionality of the latents D are determined through performance on held-out test sets from 80/20 trial splits on all experiments unless specified. To select the other hyper-parameters of GP prior length-scale κ and scale σ, we evaluated the various models on a held-out validation set. |
| Hardware Specification | Yes | 1. Synthetic HD experiment: Results were computed on a personal machine (Apple Mac Book Pro, M2 Pro chip, 32G RAM). Inference converges in a few seconds. 2. Mouse HD experiment: Results were computed on a personal machine (Apple Mac Book Pro, M2 Pro chip, 32G RAM). Inference converges in a few minutes. 3. Monkey-reaching experiment: Results were computed on an external cluster equipped with NVIDIA A100 GPUs. Fitting a single model typically used 60GB of GPU memory. All CLDS models were run strictly on CPU-only nodes (Intel Xeon Silver 4309Y Processor, 2.80 GHz, 8 cores/16 threads), using approximately 2 GB of RAM per run. Wall-clock times were between ten minutes and one hour. |
| Software Dependencies | No | The paper mentions software like 'gp SLDS', 'LFADS' (Jax implementation), and 'pynapple (PYthon Neural Analysis Package)' in Appendix B.1 and B.4, but it does not specify explicit version numbers for these software packages or the programming language (e.g., Python) used. |
| Experiment Setup | Yes | We initialize the EM algorithm at samples from our GP priors for F. With the EM algorithm we also learn the covariance parameters {Q1, Q, R}. The hyper-parameters {L, κ, σ} for the GP priors and the dimensionality of the latents D are determined through performance on held-out test sets from 80/20 trial splits on all experiments unless specified. Throughout all experiments, we ve set L = 5 to balance expressivity and number of parameters. Finally, a dimensionality of D = 2 for the head-direction experiments is chosen throughout for interpretability, as the latent space is thought to encode head-direction. For the macaque center-out reaching data, we computed co-smoothing over D {3, 5, 10, 15} (Table 2) and selected D = 5. |