niCNN: A Novel Neuromorphic Approach to Energy-Efficient and Lightweight Human Activity Recognition on Edge Devices

Recent years have seen a surge in the use of deep learning for human activity recognition (HAR) in various applications. However, running complex deep learning models on edge devices with limited resources, such as processing power, memory, and energy, is challenging. The objective of this study is to design a novel, lightweight and energy-efficient neuromorphic inspired CNN (niCNN) architecture for real-time HAR on edge devices. The niCNN architecture consists of four stages: design of a shallow CNN, conversion into an equivalent spiking network using Clamping and Quantization (CnQ) algorithm to minimize information loss, threshold balancing to calculate spiking neuron firing rate using Threshold Firing (TF) algorithm, and edge deployment. The experimental evaluation shows that the niCNN architecture achieves 97.25% and 98.92% accuracy on two publicly accessible HAR datasets, WISDM and mHealth. Furthermore, the niCNN technique retains a low inference latency of 2.25 ms and 2.36 ms, as well as a low memory utilization of 22.11 KB and 31.84 KB, respectively. Furthermore, energy usage is reduced to 5.2w and 5.8w. In comparison to various state-of-the-art and baseline CNN models, the niCNN architecture outperforms them in terms of classification metrics, memory usage, energy consumption, and inference delay. The CnQ algorithm reduces memory usage and inference latency, while the TF algorithm improves classification accuracy. The findings show that neuromorphic computing has a lot of potential for resource-constrained edge devices.