This library 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.
1. Example usage: template tasks
Norse comes packed with a few example tasks, such as MNIST, but is generally meant for use in specific deep learning tasks (see below section on long short-term spiking neural networks):
python -m norse.task.mnist
You can also run one of the jupyter notebooks on google collab.
2. Getting Started
Norse is a machine learning library that builds on the PyTorch infrastructure. While we have a few tasks included, it is meant to be used in designing and evaluating experiments involving biologically realistic neural networks.
This readme explains how to install norse and apply it in your own experiments. If you just want to try out the library, perhaps the best option is to run one of the jupyter notebooks on google collab.
Note that this guide assumes you are in a terminal friendly environment with access to the
git commands. Python version 3.7+ is required.
2.1.1. Installing from source
For now this is the recommended way of installing the package, make sure that you have installed torch, following their installation instructions and then install norse.
You can either directly install the library from github using pip:
git clone https://github.com/norse/norse cd norse pip install -e .
2.1.2. Installing from PyPi
pip install norse
2.2. Running examples
- 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
The default choices of hyperparameters are meant as reasonable starting points. More information is available when typing:
python -m norse.task.<task> --help, where
<task> is the above listed task names.
2.3. Example on using the library: 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 one interesting way to apply norse:
from norse.torch.module import LSNNLayer, LSNNCell # LSNNCell with 2 inputs and 10 outputs layer = LSNNLayer(LSNNCell, 2, 10) # 5 batch size running on CPU state = layer.initial_state(5, "cpu") # Generate data of shape [5, 2, 10] data = torch.zeros(2, 5, 2) # Tuple of output data and layer state output, new_state = layer.forward(data, state)
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.
A number of projects exist that attempts to leverage the strength of bio-inspired neural networks, however only few of them integrate with modern machine-learning libraries such as PyTorch or Tensorflow and many of them are no longer actively developed.
We are passionate about Norse and believe it has significant potential outside our own research. Primarily because we strive to follow best practices and promise to maintain this library for the simple reason that we depend on it ourselves. Second, we have implemented a number of neuron models, synapse dynamics, encoding and decoding algorithms, dataset integrations, tasks, and examples. While we are far from the comprehensive coverage of simulators used in computational neuroscience such as Brian, NEST or NEURON, we expect to close this gap as we continue to develop the library.
Finally, we are working to keep Norse as performant as possible. Preliminary benchmarks suggest that on small networks of up to ~10000 neurons Norse achieves excellent performance. We aim to create a library that scales from a single laptop to several nodes on a HPC cluster. We expect to be significantly helped in that endeavour by the preexisting investment in scalable training and inference with PyTorch.
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.
- 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). This approach maps to, but does not build on, the deep learning framework Tensorflow, which is fundamentally different from incorporating the spiking constructs into the framework itself. In turn, this requires manual translations into each individual backend, which influences portability.
- 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.
- 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.
- SpyTorch presents a set of tutorials for training SNNs with the surrogate gradient approach SuperSpike by F. Zenke, and S. Ganguli (2017). Norse implements SuperSpike, 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.
- Rockpool is a Python package developed by SynSense for training, simulating and deploying spiking neural networks. It offers both JAX and PyTorch primitives.
Please refer to the contributing.md
Norse is created by
- Christian Pehle (@GitHub cpehle), doctoral student at University of Heidelberg, Germany.
- Jens E. Pedersen (@GitHub jegp), doctoral student at KTH Royal Institute of Technology, Sweden.
LGPLv3. See LICENSE for license details.