Algorithmic Design for Neuromorphic Hardware Using Spiking Spin-Glass Models

Neural processors compute using discrete time signals, and algorithms for these machines must efficiently incorporate spiking data. We draw upon the dynamics of Hopfield networks to design a spiking neuron Potts-model for use in community detection and label propagation algorithms. Our approach to neural network design avoids the need for large training data sets to determine the hyper-parameters associated with our model, opting instead for heuristic weight-setting rules and analytical parameter setting based on the nonlinear dynamics of coupled systems of leaky-integrate and fire neurons. We have tested our approach on graphs with 128 vertices. For graphs with known ground truths, we can identify community labels sets with accuracy near 100%.