Hardware-aware Neural Architecture Search for Sound Classification in Constrained Environments

Deep Learning (DL) and Internet of Things (IoT) based applications have indeed become integral to many smart applications. Designing such solutions with high performance and low resource consumption is challenging. We present a Hardware-aware Neural Architecture Search (HW-NAS) process that automates the design of lightweight Convolutional Neural Networks (CNNs) for sound classification on edge devices. This uses a search space containing blocks similar to VGG16, a black box search strategy based on Occam's razor and an STM32 microcontroller-based simulation to estimate the hardware cost. The novel HW-NAS-derived model outperforms the existing CNN classifiers with an accuracy of 83.05% and a model size and a peak RAM consumption of 230kB and 1.22MB, respectively. Henceforth, this study concludes that HW-NAS processes can yield models with state-of-the-art performance in the domain of forest sound classification that can be deployed in resource-constrained edge devices.