Abstract

This work delves into the exploration of optimizing Multilayer Perceptrons (MLP) or the dense layers of other sorts of Deep Neural Networks when they are aimed at edge computing applications such as Internet of Things (IoT) devices, very limited in resources at the edge. The proposed optimization approach consists of generating a pruning mask for the hidden dense layers of the original neural network by using auxiliary dense Morphological Neural Networks (MNN). These MNN have shown a notable efficiency when it comes to the process of pruning, resulting in a significant decrease in the overall number of connections and a low cost in terms of accuracy degradation. The effectiveness of this new pruning methodology has been explained in detail and validated for two widely used datasets as MNIST and Fashion MNIST and two very well-known neural networks such as LeNet-5 and LeNet-300-100. Subsequently, the performance of these pruned neural networks has been assessed using an IoT hardware platform. The experimental results have outperformed other contemporary state-of-the-art pruning techniques, in terms of power efficiency and processing speed for a similar percentage of weight reduction, all while maintaining minimal impact on overall accuracy. In addition, a custom software tool has been developed to generate a C code designed to optimize the inference of these pruned networks on IoT edge devices. These findings hold important implications for advancing the development of efficient and scalable deep learning models that are specifically tailored to meet the demands of edge computing applications.