A Practical Approach to the Analysis and Optimization of Neural Networks on Embedded Systems.

The exponential increase in internet data poses several challenges to cloud systems and data centers, such as scalability, power overheads, network load, and data security. To overcome these limitations, research is focusing on the development of edge computing systems, i.e., based on a distributed computing model in which data processing occurs as close as possible to where the data are collected. Edge computing, indeed, mitigates the limitations of cloud computing, implementing artificial intelligence algorithms directly on the embedded devices enabling low latency responses without network overhead or high costs, and improving solution scalability. Today, the hardware improvements of the edge devices make them capable of performing, even if with some constraints, complex computations, such as those required by Deep Neural Networks. Nevertheless, to efficiently implement deep learning algorithms on devices with limited computing power, it is necessary to minimize the production time and to quickly identify, deploy, and, if necessary, optimize the best Neural Network solution. This study focuses on developing a universal method to identify and port the best Neural Network on an edge system, valid regardless of the device, Neural Network, and task typology. The method is based on three steps: a trade-off step to obtain the best Neural Network within different solutions under investigation; an optimization step to find the best configurations of parameters under different acceleration techniques; eventually, an explainability step using local interpretable model-agnostic explanations (LIME), which provides a global approach to quantify the goodness of the classifier decision criteria. We evaluated several MobileNets on the Fudan Shangai-Tech dataset to test the proposed approach.