A project of the Computational Physiology Laboratory at Cornell University.

Sapicore is a spiking neural network (SNN) simulator built with PyTorch. It streamlines the design and iterative testing of neuroscience-inspired models with nontrivial dynamics.

We provide programmatic and YAML-based APIs for specifying network structure and behavior, designing simulation pipelines, utilizing models for classification and clustering, and data visualization.

Sapicore's incremental class hierarchy includes efficient default implementations of neuron and synapse models commonly used in neuroscience and machine learning applications. Users may write their own dataloaders, derivative classes, and specification files in separate repositories.

Sapicore interfaces with industry-standard ML libraries, including ray, scikit-learn, networkX, and tensorboard. Object configuration using tree-config is also supported.

Sapicore 0.3.0 is in beta. The current version includes the following features:

• Spiking neurons (LIF and IZ).
• Analog neurons and oscillators.
• Static and STDP synapses.
• Network design and simulation tools.
• Scikit-compatible model API (fit/predict).
• Sampling and cross validation tools.
• Data classes with basic ETL support.
• Visualization tools.

To simulate a network from a YAML configuration using the default simulation pipeline:

python simple.py -config /path/to/config_file.yaml

See tutorials and tests/engine/network/test_network for instructive scripts and YAML files.

Basic requirements include Python 3.10+, PyTorch 1.12+, NetworkX, and the scientific stack (numpy, scipy, pandas). Developers should also install pytest and sphinx. See setup.py for more information.

To install the latest stable version of Sapicore:

pip install sapicore

To install the most recent development version:

pip install https://github.com/cplab/sapicore/archive/refs/heads/main.zip

To run tests, change directory to the one containing pytest.ini (sapicore) and call:

pytest -v -s

To run tests with a coverage report, run sapicore/tests/scripts/run_tests.py. The coverage report will be generated in a separate directory on the same level (open index.html).

If you would like to modify or extend this library:

• Clone the repository:

  git clone https://github.com/cplab/sapicore.git
  

• Create a conda virtual environment (optional):

  conda create -n <env_name> python=3.10
  conda activate <env_name>
  

• Change directory to sapicore and install with pip:

  cd sapicore
  pip install -e .
  

This project uses black to format code and flake8 for linting. We support pre-commit. To configure your local environment, install these development dependencies and set up the commit hooks:

pip install black flake8 pre-commit
pre-commit install

Documentation can be compiled by installing Sphinx and RTD, then running docs/refresh.sh.

Sapinet, the primary focus of this effort, is a multilayer spiking model designed for few-shot online learning of multiple inputs without catastrophic forgetting and without the need for data-specific hyperparameter retuning. Key features of Sapinet include denoising, regularization, scaling, classification, and stimulus similarity mapping.

If you use Sapicore, please cite the following article:

• Borthakur, A. (2022). Sapinet: A sparse event-based spatiotemporal oscillator for learning in the wild. arXiv:2204.06216.

If you are interested in dynamical perspectives on neural computation, oscillatory synchronization, or top-down control over representational state transitions, check out and cite the following articles:

• Moyal, R. & Edelman, S. (2019). Dynamic Computation in Visual Thalamocortical Networks. Entropy, 21(5).

• Edelman, S. & Moyal, R. (2017). Fundamental computational constraints on the time course of perception and action. Progress in Brain Research, 236, 121-141.

Roy Moyal, Matthew Einhorn, Jeremy Forest, Ayon Borthakur, Thomas Cleland.

• Engine, simulator, pipelines, data classes, and visualization tools by Roy Moyal.
• Framework architecture by Matthew Einhorn and Roy Moyal.
• Algorithms by Roy Moyal, Ayon Borthakur, and Thomas Cleland.
• Tutorials, utilities, and examples by Roy Moyal and Jeremy Forest.