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..
Scalable inference of functional neural connectivity at submillisecond timescales
Authors: Arina Medvedeva, Edoardo Balzani, Alex H Williams, Stephen Keeley
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Applied to both synthetic and real spike-time data from rodent hippocampus, our methods demonstrate superior accuracy and scalability compared to traditional binned GLMs, enabling functional connectivity inference in large-scale neural recordings that are temporally precise on the order of synaptic dynamical timescales and in agreement with known anatomical properties of hippocampal subregions. We validate our models on both simulated and real spiking data. In simulations, we find that both MC and PA approaches scale favorably in compute time with recording length and population size, and show improved filter recovery compared to both the discrete polynomial approximate method and traditional GLMs. We then apply our method to real spiking data, where we analyze spike-time recordings from multiple rodent hippocampal regions [22] in a dataset whose size is computationally prohibitive for traditional batched GLMs. |
| Researcher Affiliation | Collaboration | Arina Medvedeva Flatiron Institute New York, NY EMAIL Edoardo Balzani Flatiron Institute New York, NY EMAIL Alex H Williams Flatiron Institute, New York University New York, NY EMAIL Stephen L Keeley Fordham University New York, NY EMAIL |
| Pseudocode | No | The paper describes methods and derivations using mathematical equations and prose but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide open-source implementations of both MC and PA estimators, optimized for GPU acceleration, to facilitate adoption in the neuroscience community1. 1The Poisson point process GLM code is available at https://github.com/macari216/ poisson-process-glm.git |
| Open Datasets | Yes | We use publicly available data from the Allen Institute consisting of 106 neurons (NCA1 = 62, NCA3 = 28, NDG = 16) recorded with a single probe over approximately 2.7 hours [22]. [22] Allen Institute for Brain Science. Visual coding neuropixels, 2023. URL https://portal.brain-map. org/explore/circuits/visual-coding-neuropixels. Dataset includes spike times, LFP, and behavior from mouse visual cortex during stimuli presentation. |
| Dataset Splits | No | The paper mentions using "simulated data" and "real spiking data" and that for real data, filter accuracy was assessed on "various subsets of the data". However, it does not provide specific details on how these subsets were formed (e.g., exact percentages for train/test/validation splits, explicit sample counts, or defined cross-validation strategies). It notes that CCGs were calculated from the "full dataset" for comparison, but the process of splitting the data for training and evaluation is not explicitly described. |
| Hardware Specification | No | The abstract mentions "optimized for GPU acceleration", and the methods state "continuous-time methods utilize GPU-parallelized scans over the data". However, the paper does not specify any particular GPU models (e.g., NVIDIA A100, Tesla V100), CPU models, or memory details used for the experiments. It only refers to GPUs in a general sense. |
| Software Dependencies | No | The paper mentions software components like "SVRG optimizer" and discusses "Matplotlib" for plotting in figure captions. It also implies the use of standard scientific computing libraries, but no specific version numbers for any software, libraries, or frameworks (e.g., Python 3.x, PyTorch 1.x) are provided in the main text. |
| Experiment Setup | Yes | In this work, we select history window length H of 4-6 ms to encompass expected neuronal dynamical effects. In our analyses of neural recordings, we use an approximation interval spanning 3-7 Hz around the mean rate. We set c = 1.5 and α = 2 throughout the manuscript based on initial model exploration. We run all simulations with 100 RC bases and fit all models with 3 to 5 GL bases. All models are run with ridge regularization (β = 1000), a common choice for GLMs [3, 4, 9], to encourage sparsity in synaptic connections (see Supplement S.1 for more hyperparameter details). For discrete models, the bin size is set to 0.1 ms. |