Draft:Norse (neuron simulator)

• Comment: Maybe it's too early for the subject to be notable, or more independent reliable sources needed to show notability. Artem.G (talk) 18:47, 28 March 2022 (UTC)

---

Norse
Initial release	December 2019
Repository	https://github.com/norse/norse
Written in	Python
License	LGPLv3
Website	https://norse.ai

Norse is a simulator for biological neuron models with an emphasis on gradient-based optimization and integration with neuromorphic hardware and event-based cameras. Norse is developed primarily by researchers at Heidelberg University and KTH Royal Institute of Technology and is publicly available under the LGPLv3 license.

Differences from other simulators

Norse separates itself from the more "classical" branch of simulators like Neuron (software), NEST (software), and Brian (software) in that it models neural networks as directed graphs that purely operate on tensors, similar to deep learning libraries like PyTorch and TensorFlow. Such computational graphs lend themselves well to hardware acceleration on CPUs, GPUs, and TPUs. It also enables automatic differentiation for neuron discontinuities, such as spiking neural network via gradient approximation methods like SuperSpike, which Norse implements to allow the training of arbitrarily deep networks using backpropagation through time.

Example

The following example demonstrates a network that combines leaky integrate-and-fire neurons with conventional convolutions to classify digits from the MNIST dataset with >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)

See also

• Biological neuron model
• Neuromorphic engineering
• Spiking neural network

External references

• Public code repository on GitHub: https://github.com/norse/norse/

Category:Simulation software Category:Computational neuroscience