Simulation of a Novel, Low Swap, Sparse Hyper-Dimensional Neural Network Architecture for Anomaly Detection AI AT the Edge

This paper details the simulation and performance results of a Sparse Hyper-Distributed Robust Efficient Neural Network (SpHyRE-Net) architecture that performs anomaly detection for real-world time-series data. SpHyRE-Net is an innovative, novel, low size, weight and power (SWaP) machine learning solution for devices operating at the tactical edge. It utilizes bit operations and sparse hyper-dimensional representations for bio-inspired learning via a Hebbian-like rule that results in a combined power-latency reduction of 2-orders of magnitude over ordinary deep networks. The paper details the application of SpHyRE-Net to real-world cell-traffic datasets as well as the simulation requirements needed to minimize latency and memory use. Also discussed are the mechanisms necessary for implementing the architecture on an FPGA as a precursor to realization on a neuro-morphic ASIC with ultra-low power profile.