MultiBarrelModel

Code implementation of our paper 'Localist Topographic Expert Routing: A Barrel Cortex-Inspired Modular Network for Sensorimotor Processing' (NeurIPS 2025).

Dependency

Core dependencies: Python 3.10 and PyTorch 1.12.1. See requirements.txt for additional packages. Run

pip install -r requirements.txt

Datasets

This work primarily uses the EvTouch-Objects and EvTouch-Containers tactile datasets, consisting of 36 and 20 classes respectively. For detailed information about these datasets, please refer to TactileSGNet.

File description

• /data/: Contains raw files for both tactile datasets (EvTouch-Objects and EvTouch-Containers).
• dataset.py: Loads two tactile datasets.
• MultiBarrel4EvTask.py: Train a multi-barrel model with 39 independently parameterized barrels.
• SharedMultiBarrel4EvTask.py: Train a multi-barrel model with 39 barrels sharing training parameters.
• SingleBarrel4EvTask.py: Train a single-barrel model with neuron count matching the above two models.
• utils.py: Some auxiliary modules in the model (e.g., single-neuron dynamics).
• MultiBarrel_simulate.py: Simulate optogenetic experiments to observe the spread of neural activity.
• Model_lossLandscape.py: Visualize the loss landscape of models.
• MultiBarrel_propagation.py: Measure neural activity correlation between barrels.

Train models

The code is almost one-click runnable. Once the dataset files are correctly placed in the ./data/ directory, you can train either shared-parameter or independent-parameter multi-barrel models by executing the corresponding .py file directly. For example:

python SharedMultiBarrel4EvTask.py

Note that the EvTouch-Objects and EvTouch-Containers datasets contain different numbers of classes.

Simulate optogenetic experiments

Similarly, run MultiBarrel_simulate.py to visualize the temporal spread of neural activity:

python MultiBarrel_simulate.py

Loss landscape

Load the trained model weights to visualize the loss landscape:

python Model_lossLandscape.py

Measure neural activity correlation

Load the trained model weights to calculate both global and local neural activity correlations across barrel pairs:

python MulitBarrel_Propagation.py

Citation

If you find this work useful, please cite:

@inproceedings{  
2025localist,  
title={Localist Topographic Expert Routing: A Barrel Cortex-Inspired Modular Network for Sensorimotor Processing},  
author={Tianfang Zhu and Dongli Hu and Jiandong Zhou and Kai Du and Anan LI},  
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},  
year={2025},  
url={https://openreview.net/forum?id=1Y8MXuJlIY}  
}

Feel free to raise any questions related to this work.