NLP4Neuro

Off-the-shelf, pre-trained large language models applied to neural population decoding.

Getting started

1. Download the data from here [https://doi.org/10.7910/DVN/I8LULX] into the "data" directory.

From a GNU terminal window:

wget -r --no-parent data "https://dataverse.harvard.edu/api/datasets/:persistentId/dirindex?persistentId=doi:10.7910/DVN/I8LULX&folder=exp1-4_data"

From Windows Powershell:

Invoke-WebRequest -Uri "https://dataverse.harvard.edu/api/datasets/:persistentId/dirindex?persistentId=doi:10.7910/DVN/I8LULX&folder=exp1-4_data" -OutFile "exp1-4_data"

Alternatively, there is a sample download_dataverse powershell script provided in the "data" folder.

2. Update the config.yaml file with the locations of your data and results folders.

3. Activate a conda environment corresponding to the experiment you wish to run.

user@remotepc ~/nlp4neuro/code/experiment_1 bash create_jjm_env_nlp1.sh

4. Run the experiment, or view experiment results, found in the "plot_results" folder.

(jjm_env_nlp1) user@remotepc ~/nlp4neuro/code/experiment_1 bash run_experiment1.sh

(jjm_env_nlp3) user@remotepc ~/nlp4neuro/code/plot_results python exp3_create_pcaplots.py

Conda environments

• Each experiment provides a .yml file and script for activating the corresponding environment.

Assigning a GPU

• All jobs were run using the Duke Compute Cluster (DCC) in conjunction with SLURM workload manager.
• E.g. in run_experiment_1.sh, we ran jobs requiring 64 GB of RAM, 8 CPUs per task, and an A6000 GPU

#!/bin/bash
#SBATCH --job-name=nlp4neuro          # Job name
#SBATCH --output=logs/nlp4neuro_exp1_%j.out # Standard output file (%j will be replaced by the job ID)
#SBATCH --error=logs/nlp4neuro_exp1_%j.err  # Standard error file
#SBATCH --time=72:00:00                    # Max run time (adjust as needed)
#SBATCH --mem=64G                         # Memory request per node
#SBATCH --cpus-per-task=8                  # Number of CPU cores per task
#SBATCH --mail-type=END,FAIL               # Email notifications for job done & fail
#SBATCH [email protected] # Your email for notifications
#SBATCH -p scavenger-gpu --gres=gpu:6000_ada:1 #Select the GPU

• Please note that each experiment .sh script header data corresponds to this particular setup.
• You may alter each .sh script based on your work environment, or simply run the experiment .py file.

(my_env) user@remotepc ~/nlp4neuro/code/experiment_1 python run_experiment1.py

Web links for code/data

• NLP4Neuro code: [https://github.com/Naumann-Lab/nlp4neuro]
• NLP-4-Neuro data: [https://doi.org/10.7910/DVN/I8LULX]

Systems/envs used for testing

1. Lenovo Legion Slim 5 14APH8, 32 GB RAM, NVidia RTX 4060 8 GB, Windows 11 Home
2. Duke Compute Cluster (DCC) remote access, NVidia A6000 via Windows Powershell

Documentation written by Jacob Morra. For clarification you may contact me at [[email protected]].

Citing this work

If you wish to use this platform or any of the provided data/code, please cite our paper:

J. J. Morra et al., “NLP4Neuro: Sequence-to-sequence learning for neural population decoding,” arXiv [q-bio.NC], 03-Jul-2025.