isingToSNN

Background

Within this project the mapping process of a generic ISING model described by the Hamiltonian:

$$H=-\sum_i^N \sum_j^N J_{ij} s_i s_j-\sum_i h_i s_i$$;

in the field of Spiking Neural Networks.

The picture below schematically summarizes the different steps:

The pipeline foreseen for the implementation consists of a series of steps, first of all the process of mapping the objective function $f(\textbf{x})$ into the Hamiltonian of ISING $H(\textbf{s})$:

Once obtained the map of the couplings $J$, and the external magnetic field $h$, through the mapping between the QUBO to ISING. In order to map the corresponding Hamiltonian, reference is made to the mapping process described by Fonseca et al.

A first network prototype is then developed through the GeNN platform, a GPU-enhanced neural network simulation environment based on code generation for Nvidia CUDA, which allows to perform numerous simulations in a short time. Furthermore, in order to improve the performance of the network in the search for the optimum, the tool NNI is used which allows to provide an optimization of the hyperparameters associated with the network.

In order to evaluate the actual capabilities of the network, a code deployment will be performed on SpiNNaker 1 and SpiNNaker 2, again through a hyperparameter optimization process in order to compensate for the weight quantization process.

Reference

The reference material used to develop the research is obtained from:

• Fonseca Guerra, Gabriel A., and Steve B. Furber. "Using stochastic spiking neural networks on spinnaker to solve constraint satisfaction problems." Frontiers in neuroscience 11 (2017): 714.
• Knight, James C., Anton Komissarov, and Thomas Nowotny. "PyGeNN: a Python library for GPU-enhanced neural networks." Frontiers in Neuroinformatics 15 (2021): 659005.

Virtual environment configuration

The project makes use of the virtual environment creator conda.

Conda installation:

wget https://repo.anaconda.com/miniconda/Miniconda3-py38_4.10.3-Linux-x86_64.sh
chmod 777 Miniconda3-py38_4.10.3-Linux-x86_64.sh
./Miniconda3-py38_4.10.3-Linux-x86_64.sh

Virtual environment creation:

conda create --name cspPy36 python=3.6.13

Environment activation

conda activate cspPy36

Package installation, through the following commands the package numpy and notebook(Jupyter Notebook) will be installed, the version of PyGeNN that is recommended to install can be found at the link

conda install pip
pip install numpy notebook
pip install pygenn-4.7.1-cp36-cp36m-linux_x86_64.whl
python -m ipykernel install --user --name=cspPy36

Project Structure

hyperdimNearestNeighbors: contains the script that allows you to compute the first neighbors, for a regular lattice with an arbitrary number of spins (per dimension) in n-dimensions; bruteForceSolver: returns the minimum energy value and the corresponding configurations, using the brute-force method; isingToSNN: main script which contains the description of the problem to be solved, the mapping in SNN and the neuron activity plot.

Contact

• riccardo.pignari@polito.it
• gianvito.urgese@polito.it