Illustration to the Differentiable Approach for Multi-scale Brain Modeling

Here, we illustrate to implement a differentiable approach for multi-scale brain modeling. See our paper for more details:

• C. Wang, M. Lyu, T. Zhang, and S. Wu, “A differentiable approach to multi-scale brain modeling,” in ICML 2024 Workshop on Differentiable Almost Everything, 2024.

Requirements

In general, several Python packages related to the brainpy ecosystem are required to run the code:

• brainstate
• brainunit
• brainscale
• braintools
• brainpy-datasets

Neuron fitting

We provide a simple example to illustrate the differentiable neuron fitting process.

• To fit the GIF model based on the membrane potential or spike trains, run the following command:

python neuron_fitting_of_gif_model.py

• To fit the HH model based on the membrane potential or spike trains, run the following command:

python neuron_fitting_of_hh_model.py

Training conductance-based EI spiking networks on cognitive tasks

We provide a simple example to illustrate the training process of conductance-based EI spiking networks based on the cognitive task of Evidence Accumulation.

The following arguments are used to control the model configuration and training:

• --conn_method: The method to generate the connection matrix. It can be dense (dense connection), gaussian (sparse connection), or rand (random and sparse connection).
• --n_rec: The number of recurrent neurons. Then, the excitatory and inhibitory neurons are divided by a 4:1 ratio.
• --w_ei_ratio: The E/I weight ratio.
• --mode: It can be train (train the network) or sim (simulate the network).
• --method: The training method. Can be bptt (back-propagation through time) or expsm_diag or diag (if brainscale is available).

For example, to train a conductance-based EI spiking network, run the following command:

# Training with BPTT
python task-coba-ei-rsnn.py --method bptt

# Training with online learning methods in BrainScale
python task-coba-ei-rsnn.py --method diag
python task-coba-ei-rsnn.py --method expsm_diag --etrace_decay 0.98