electronicvisions / norse

A deep learning library for spiking neural networks.

Norse aims to exploit the advantages of bio-inspired neural components, which are sparse and event-driven - a fundamental difference from artificial neural networks. Norse expands PyTorch with primitives for bio-inspired neural components, bringing you two advantages: a modern and proven infrastructure based on PyTorch and deep learning-compatible spiking neural network components.

Documentation: norse.github.io/norse/

1. Getting started

The fastest way to try Norse is via the jupyter notebooks on Google collab.

Alternatively, you can install Norse locally and run one of the included tasks such as MNIST:

python -m norse.task.mnist

2. Using Norse

Norse presents plug-and-play components for deep learning with spiking neural networks. Here, we describe how to install Norse and start to apply it in your own work. Read more in our documentation.

2.1. Installation

We assume you are using Python version 3.7+ and have installed PyTorch version 1.9 or higher. Read more about the prerequisites in our documentation.

pip install norse

pip install -qU git+https://github.com/norse/norse

docker pull quay.io/norse/norse

conda install -c norse norse

For troubleshooting, please refer to our installation guide, create an issue on GitHub or write us on Discord.

2.2. Running examples

Norse is bundled with a number of example tasks, serving as short, self contained, correct examples (SSCCE). They can be run by invoking the norse module from the base directory. More information and tasks are available in our documentation and in your console by typing: python -m norse.task.<task> --help, where <task> is one of the task names.

• To train an MNIST classification network, invoke

  python -m norse.task.mnist

• To train a CIFAR classification network, invoke

  python -m norse.task.cifar10

• To train the cartpole balancing task with Policy gradient, invoke

  python -m norse.task.cartpole

Norse is compatible with PyTorch Lightning, as demonstrated in the PyTorch Lightning MNIST task variant (requires PyTorch lightning):

python -m norse.task.mnist_pl --gpus=4

2.3. Example: Spiking convolutional classifier

This classifier is a taken from our tutorial on training a spiking MNIST classifier and achieves >99% accuracy.

import torch, torch.nn as nn
from norse.torch import LICell             # Leaky integrator
from norse.torch import LIFCell            # Leaky integrate-and-fire
from norse.torch import SequentialState    # Stateful sequential layers

model = SequentialState(
    nn.Conv2d(1, 20, 5, 1),      # Convolve from 1 -> 20 channels
    LIFCell(),                   # Spiking activation layer
    nn.MaxPool2d(2, 2),
    nn.Conv2d(20, 50, 5, 1),     # Convolve from 20 -> 50 channels
    LIFCell(),
    nn.MaxPool2d(2, 2),
    nn.Flatten(),                # Flatten to 800 units
    nn.Linear(800, 10),
    LICell(),                    # Non-spiking integrator layer
)

data = torch.randn(8, 1, 28, 28) # 8 batches, 1 channel, 28x28 pixels
output, state = model(data)      # Provides a tuple (tensor (8, 10), neuron state)

2.4. Example: Long short-term spiking neural networks

The long short-term spiking neural networks from the paper by G. Bellec, D. Salaj, A. Subramoney, R. Legenstein, and W. Maass (2018) is another interesting way to apply norse:

import torch
from norse.torch import LSNNRecurrent
# Recurrent LSNN network with 2 input neurons and 10 output neurons
layer = LSNNRecurrent(2, 10)
# Generate data: 20 timesteps with 8 datapoints per batch for 2 neurons
data  = torch.zeros(20, 8, 2)
# Tuple of (output spikes of shape (20, 8, 2), layer state)
output, new_state = layer(data)

3. Why Norse?

Norse was created for two reasons: to 1) apply findings from decades of research in practical settings and to 2) accelerate our own research within bio-inspired learning.

We are passionate about Norse: we strive to follow best practices and promise to maintain this library for the simple reason that we depend on it ourselves. We have implemented a number of neuron models, synapse dynamics, encoding and decoding algorithms, dataset integrations, tasks, and examples. Combined with the PyTorch infrastructure and our high coding standards, we have found Norse to be an excellent tool for modelling scaleable experiments and Norse is actively being used in research.

Finally, we are working to keep Norse as performant as possible. Preliminary benchmarks suggest that Norse achieves excellent performance on small networks of up to ~5000 neurons per layer. Aided by the preexisting investment in scalable training and inference with PyTorch, Norse scales from a single laptop to several nodes on an HPC cluster with little effort. As illustrated by our PyTorch Lightning example task.

Read more about Norse in our documentation.

4. Similar work

The list of projects below serves to illustrate the state of the art, while explaining our own incentives to create and use norse.

• BindsNET also builds on PyTorch and is explicitly targeted at machine learning tasks. It implements a Network abstraction with the typical 'node' and 'connection' notions common in spiking neural network simulators like nest.
• cuSNN is a C++ GPU-accelerated simulator for large-scale networks. The library focuses on CUDA and includes spike-time dependent plasicity (STDP) learning rules.
• decolle implements an online learning algorithm described in the paper "Synaptic Plasticity Dynamics for Deep Continuous Local Learning (DECOLLE)" by J. Kaiser, M. Mostafa and E. Neftci.
• GeNN compiles SNN network models to NVIDIA CUDA to achieve high-performing SNN model simulations.
• Long short-term memory Spiking Neural Networks (LSNN) is a tool from the University of Graaz for modelling LSNN cells in Tensorflow. The library focuses on a single neuron and gradient model.
• Nengo is a neuron simulator, and Nengo-DL is a deep learning network simulator that optimised spike-based neural networks based on an approximation method suggested by Hunsberger and Eliasmith (2016).
• Nengo PyTorch a thin wrapper for PyTorch that adds a single voltage-only spiking model. The approach is independent from the Nengo framework.
• Neuron Simulation Toolkit (NEST) constructs and evaluates highly detailed simulations of spiking neural networks. This is useful in a medical/biological sense but maps poorly to large datasets and deep learning.
• PyNN is a Python interface that allows you to define and simulate spiking neural network models on different backends (both software simulators and neuromorphic hardware). It does not currently provide mechanisms for optimisation or arbitrary synaptic plasticity.
• PySNN is a PyTorch extension similar to Norse. Its approach to model building is slightly different than Norse in that the neurons are stateful.
• Rockpool is a Python package developed by SynSense for training, simulating and deploying spiking neural networks. It offers both JAX and PyTorch primitives.
• Sinabs is a PyTorch extension by SynSense. It mainly focuses on convolutions and translation to neuromorphic hardware.
• SlayerPyTorch is a Spike LAYer Error Reassignment library, that focuses on solutions for the temporal credit problem of spiking neurons and a probabilistic approach to backpropagation errors. It includes support for the Loihi chip.
• SNN toolbox automates the conversion of pre-trained analog to spiking neural networks. The tool is solely for already trained networks and omits the (possibly platform specific) training.
• snnTorch is a simulator built on PyTorch, featuring several introduction tutorials on deep learning with SNNs.
• SpikingJelly is another PyTorch-based spiking neural network simulator. SpikingJelly uses stateful neurons. Example of training a network on MNIST.
• SpyTorch presents a set of tutorials for training SNNs with the surrogate gradient approach SuperSpike by F. Zenke, and S. Ganguli (2017). Norse implements the SuperSpike surrogate gradient function, but allows for other surrogate gradients and training approaches.
• s2net is based on the implementation presented in SpyTorch, but implements convolutional layers as well. It also contains a demonstration how to use those primitives to train a model on the Google Speech Commands dataset.

5. Contributing

Contributions are warmly encouraged and always welcome. However, we also have high expectations around the code base so if you wish to contribute, please refer to our contribution guidelines.

6. Credits

Norse is created by

• Christian Pehle (@GitHub cpehle), PostDoc at University of Heidelberg, Germany.
• Jens E. Pedersen (@GitHub jegp), doctoral student at KTH Royal Institute of Technology, Sweden.

More information about Norse can be found in our documentation. The research has received funding from the EC Horizon 2020 Framework Programme under Grant Agreements 785907 and 945539 (HBP) and by the Deutsche Forschungsgemeinschaft (DFG, German Research Fundation) under Germany's Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster).

7. Citation

If you use Norse in your work, please cite it as follows:

@software{norse2021,
  author       = {Pehle, Christian and
                  Pedersen, Jens Egholm},
  title        = {{Norse -  A deep learning library for spiking 
                   neural networks}},
  month        = jan,
  year         = 2021,
  note         = {Documentation: https://norse.ai/docs/},
  publisher    = {Zenodo},
  version      = {0.0.7},
  doi          = {10.5281/zenodo.4422025},
  url          = {https://doi.org/10.5281/zenodo.4422025}
}

Norse is actively applied and cited in the literature. We are keeping track of the papers cited by Norse in our documentation.

8. License

LGPLv3. See LICENSE for license details.