Rescuing neural spike train models from bad MLE

Authors: Diego Arribas, Yuan Zhao, Il Memming Park

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments performed on both real and synthetic neural data validate the proposed approach, showing that it leads to well-behaving models.
Researcher Affiliation Academia 1Department of Neurobiology and Behavior Center for Neural Circuit Dynamics Stony Brook University, NY, USA 2 Biomedicine Research Institute of Buenos Aires, Argentina
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes The following link contains the code used to fit the models https://github.com/diegoarri91/mmd-glm.
Open Datasets Yes We used two small datasets from monkey ventral premotor cortex (Monkey PMv Figs 3A-H) and human neocortex (Human Cortex Figs. 3I-P) that are prone to yield unstable ML parameters [18, 19]. ... We used a dataset recorded from the lateral intraparietal (LIP) area of a monkey during a perceptual decisionmaking task [17, 21].
Dataset Splits Yes We used 50 trials for training the models and 50 for validating.
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No The paper does not list any specific software dependencies (e.g., library or solver names with version numbers) required to replicate the experiments.
Experiment Setup Yes To determine the weight of the MMD term (α) we tried values on a grid and used the smallest α for which the MMD-GLM samples matched the data firing rate within a 10% interval. To study the variability of the stochastic optimization, we repeated the procedure 20 times and report average values. ... We initialized the coefficients of the history filter at zero and the bias at its MLE value for every optimization. We then minimized NLL + αMMD drawing 100 trials from the model at each optimization step to compute MMD and its gradient.