Markov Chain Monte Carlo and Neural Sampling Inference for Graphical Models at the Edge using the SpiNNaker2 Neuromorphic Chip-Multiprocessor

In this paper, we present results of designing and mapping approximate probabilistic inference algorithms on a 152 core ARM M4F multiprocessor with a neuromorphic event based NOC, the SpiNNaker2. Our approach implements neural sampling neuromorphic sampling algorithm well suited to the SpiNNaker2's event based (spike) processing architecture. We perform a comparative study of approximate inference on the SpiNNaker1 and the SpiNNaker2 architecture. The implemented algorithms employ a binary Bayesian tree-based algorithm as a benchmark for the task at hand. We model and characterize the networks of spiking neurons, taking advantage the hardware's event-handling capabilities while at the same time exploiting the massively parallel nature of SpiNNaker architecture.