Task-driven neural network models predict neural dynamics of proprioception: Neural network model weights

#############

Task-driven neural network models predict neural dynamics of proprioception, Cell 2024

#############

Authors: Marin Vargas, Alessandro (orcid=0000-0001-7073-4120) and Bisi, Axel (orcid=0009-0006-8602-7555) and Chiappa, Alberto Silvio (orcid=0009-0001-2764-6552) and Versteeg, Christopher (orcid=0000-0002-4269-5109) and Miller, Lee E. (orcid=0000-0001-8675-7140) and Mathis, Alexander (orcid=0000-0002-3777-2202)

Affiliation: EPFL

Date: January, 2024

Link to the Cell article: 

https://www.cell.com/cell/pdf/S0092-8674(24)00239-3.pdf

--------------------------------

Here we provide the trained model checkpoints for all tasks of our article "Task-driven neural network models predict neural dynamics of proprioception". It contains 300 temporal convolutional networks (TCNs) and 50 LSTM models trained on 16 tasks as well as the untrained initialization. 

The overall structure of the data is:

└── models
    ├── deepdraw_models             - Contains networks hyperparameters
    │   ├── template_models         - Contains the default parameters
    │   ├── torque                            - Contains network hyperparameters for the torque task
    │   └── ...                     
    ├── experiment_***                  - Contains checkpoint of trained and untrained models 
    ├── ...                         
    └── ...       

--------------------------------

The checkpoints are stored in experiment folders (experiment_***) that follow this scheme:
- Task:         shallow exp id,     deep TCNs exp id,     LSTM id.

Experiment IDs for each task:

- Untrained:                                                               15,         115,         45
- Classification:                                                        4015,   5015,   4045

- Torque:                                                                    8015,   8030,   8045

- Regress joint pos:                                                 17016,  17031,  17046
- Regress joint vel:                                                   17216,  17231,  17246
- Regress joint pos & vel::                                       17416,  17431,  17446
- Regress joint pos & vel & acc::                            20516,  20531,  20546

- Regress hand pos:                                                 4016,   5016,   4046
- Regress hand vel:                                                   17316,  17331,  17346
- Regress hand pos & vel:                                        17516,  17531,  17546
- Regress hand pos & vel & acc:                             20416,  17831,  17846

- Regress hand and elbow pos:                              20016,  20031,  20046
- Regress hand and elbow vel:                                20916,  20931,  20946
- Regress hand and elbow pos & vel:                     20616,  20631,  20646
- Regress hand and elbow pos & vel & acc:          20816,  20831,  20846

- Redundancy reduction - task transfer (AR):       10020,  10035,  10050
- Redundancy reduction - task transfer (HP):       10021,  10036,  10051
- Autoencoder                                                           20716 & 20717, 20731 & 20732,   X

The code to load, evaluate and train the models is available at: https://github.com/amathislab/Task-driven-Proprioception/tree/master/nn-training

--------------------------------

The datasets, weights, activations and predictions are released with Creative Commons Attribution 4.0 license.

The code is released under the MIT license, see https://github.com/amathislab/Task-driven-Proprioception

If you find our code, weights, predictions or ideas useful, please cite:

@article{vargas2024task,
  title={Task-driven neural network models predict neural dynamics of proprioception},
  author={{Marin Vargas}, Alessandro and Bisi, Axel and Chiappa, Alberto S and Versteeg, Chris and Miller, Lee E and Mathis, Alexander},
  journal={Cell},
  year={2024},
  publisher={Elsevier}
}