A PYNQ Z2 - NeuroShield Architecture for a Fast, Low-Energy and Privacy Preserving Image Classification

This paper presents an embedded platform for local image classification using the NM500 neuromorphic chip on a NeuroShield connected to a PYNQ-Z2 board. The system performs fully on-device training and inference on the MNIST dataset, achieving 95.17% accuracy with only 4,032 hardware neurons compared to the$\sim 8,000$used in the reference benchmark. Operating autonomously without a host processor, it delivers real-time performance with a 7.46 ms inference latency$(134 \text{img} / \mathrm{s})$and consumes under 2 W. These results show that compact NM500-based designs can provide accurate, low-power, and fully embedded intelligence for edge applications.