One-hot Generalized Linear Model for Switching Brain State Discovery

Authors: Chengrui Li, Soon Ho Kim, Chris Rodgers, Hannah Choi, Anqi Wu

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTAL EVALUATION Models for comparison. We will compare our methods and state-of-the-art baseline methods on one simulated data and two real neural datasets
Researcher Affiliation Academia 1 School of Computational Science & Engineering Georgia Institute of Technology, Atlanta, GA 30305, USA {cnlichengrui,anqiwu}@gatech.edu 2 School of Mathematics Georgia Institute of Technology, Atlanta, GA 30305, USA {soonhokim,hannahch}@gatech.edu 3 Department of Neurosurgery, School of Medicine Emory University, Atlanta, GA 30322, USA christopher.rodgers@emory.edu
Pseudocode No The paper describes inference algorithms mathematically but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/JerrySoybean/onehot-hmmglm.
Open Datasets Yes PFC-6 dataset (Peyrache et al., 2018; 2009)1. Neural spike trains were collected while a rat learned a behavioral contingency task. During recording, the animal performed a trial for about 4 secs and then took a short break for about 24 secs. The spike train data used for learning and testing is segmented from the long session. Each sequence starts from 5 seconds before a behavior start and lasts for 10 seconds after the start. Hence, each sequence corresponds to a behavioral trial. We use 2/3 of the neural sequences as the training set and the remaining 1/3 as the test set. The neural spikes are binned into 750 time bins with bin size = 20 ms.
Dataset Splits No For each trial, we train different models on the first 10 sequences and test on the remaining 10 sequences. (Simulated data) and We use 2/3 of the neural sequences as the training set and the remaining 1/3 as the test set. (PFC-6 dataset) and The first 30 trials are used in the analysis of each session, of which 10 randomly selected trials form the test set when evaluating the test log-likelihood, and the remaining 20 trials are used for training the model. (Barrel Cortex data). No explicit validation split information is provided.
Hardware Specification No The paper does not provide specific hardware details such as CPU/GPU models, memory specifications, or cloud computing resources used for experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) that were used for the experiments.
Experiment Setup Yes In our model, τ is used to force the soft one-hot close to one corner of the simplex, so we tried τ ∈ {0.1, 0.2, 0.5}, and found that the result of the one-hot HMM-GLM is not sensitive to τ in this range. Given that the selection of τ is insensitive to different datasets, we fix τ = 0.2, which is a common moderate choice. (2) Generative hyperparameters {µw, σ2w, µb, σ2b}: we chose µw = 5, σw = 2 and µb = 0, σb = 2 since this set provides noninformative priors for the weight strength and the background intensity in GLMs, and hence the inference is insensitive to different datasets.