Pruning dense neural networks for efficient edge deployment in internet of things applications

This study presents an innovative approach to optimizing Multi-Layer Perceptrons (MLPs) for resource-constrained edge computing applications, particularly Internet of Things (IoT) devices. The proposed strategy leverages auxiliary dense Morphological Neural Networks (MNNs) to generate pruning masks for the dense hidden layers of the neural network. MNNs have proven to be highly effective in eliminating redundant weights and connections, resulting in minimal accuracy loss. The efficacy of this pruning method was tested and validated using two benchmark datasets: handwritten MNIST and fashion MNIST. Subsequently, the performance of the pruned neural networks was assessed on an IoT hardware platform. Experimental results demonstrated that pruned models outperformed other contemporary pruning techniques in terms of power efficiency and processing speed, without compromising accuracy or weight reduction. Additionally, a custom software tool was developed to enhance the inference capabilities of these pruned networks on IoT edge devices. These findings have significant implications for the development of efficient and scalable deep learning models tailored to the demands of edge computing applications.