Algorithm-hardware co-design of neuromorphic networks with dual memory pathways

Pengfei Sun* [1], Zhe Su*[2], Jascha Achterberg [3], Giacomo Indiveri [2] , Dan F.M. Goodman [1], Danyal Akarca [1, 4]

Imperial College London, ETH Zurich, University of Oxford, University of Cambridge

• These authors contributed equally to this work.

Algorithm: Pengfei Sun

Hardware: Zhe Su

README

Requirements

Python 3 with the following packages installed:

• PyTorch
• numpy
• spikingjelly

The software has been tested with CUDA libraries version 11.3 and Pytorch 1.12.1, and spikingjelly==0.0.0.0.14.

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 -f https://download.pytorch.org/whl/cu113/torch_stable.html

pip install spikingjelly

Examples

The core module is "dual_memory_pathways.py", and you can call "ConvLMU2" function.

output = convLMU2(input_dim, hidden_dim, out_dim, d, TT, t)

input_dim: input dimension

hidden_dim: hidden dimension

out_dim: output dimension

d: memory state

TT: total simulation time step

t: state buffer length

Example implementations can be found inside different dataset folders. For the SHD/SSC, please download the dataset from (https://zenkelab.org/datasets/).

• Run example SHD implementation,

   cd shd/src

   python train_spiking.py -d 5 -t 40  

d is the memory state and t is the state buffer length.