邊緣裝置上的高效能隨時分類

Abstract

近年來隨著物聯網、自駕車等應用的發展，邊緣計算越來越為重要。而這些技術的發展與神經網路有著重要的關聯，許多應用都將使用到神經網路。我們發現在邊緣裝置上可能會有計算能力較低的特性，也可能在能源上有所限制或是透過外部的其他資源來獲得能量，且有許多邊緣裝置都需要使用在即時的應用中。因此我們提出了幾項方法來針對神經網路的輸入做處理，以大量減少神經網路所需的計算量，並嘗試保持神經網路的準確率。我們針對我們的方法來搭配資料集做實驗，接著對實驗結果做分析與討論，最後在邊緣裝置的使用上提出相關的建議。在三個資料集的實驗中，我們最好的方法至少可以減少網路架構50%以上的FLOPs。表明我們的方法將可以依照邊緣裝置的情境來解決邊緣裝置計算能力低、能源有限、即時應用等問題。

In recent years, with the development of applications of the Internet of Things, Self-driving cars, and other fields, Edge Computing has become increasingly important. The development of these technologies has an important relationship with neural networks, and many applications will use neural networks. We found that edge devices may have lower computing capability, or may have limited resources or obtain energy through other external resources, and many edge devices need to be used in real-time applications. Therefore, we propose several methods to process the input of the neural network to reduce the amount of calculation required by the neural network and try to maintain the accuracy of the neural network. We will experiment with the dataset according to our method, then analyze and discuss the experimental results, and finally put forward relevant suggestions on the use of edge devices. In the experiment of the three datasets, our best method can reduce the FLOPs of the network architecture by at least 50%. It shows that our method can solve the problems of low calculation capacity, limited energy, and real-time applications of edge devices according to the situation of edge devices