Research About Energy Efficiency Optimization Of CNN Models On Android Platform

In recent years,Convolutional Neural Networks(CNNs)have been widely used in various domains,such as computer vision and speech recognition because of their high accuracy and strong self-adaptiveness.On the other hand,mobile phones have become carry-ons to human beings,and generate a large number of sensor data every day.In or-der to make mobile phones serve people more intelligently,many engineering projects also try to use CNNs to process these sensor data on the mobile phone.However,due to resource limitations(memory,computing power,battery capacity,etc.)of current mo-bile platforms,CNN-based mobile applications have not become mainstream on mobile platforms.Currently,most CNN-based mobile applications adopt the client-server comput-ing paradigm.But this paradigm not only depends on network performance(such as network stability)but also leads to privacy leakage.Therefore,many researchers have begun to explore how to perform the inference process of convolutional neural networks directly on the mobile platform.For this research topic,this paper proposes a series of optimization strategies and designs a CNN inference library,which can run on the An-droid platform in an energy efficient way.Then,this paper uses the inference library to develop a life-logging application to explore ways to further improve the energy effi-ciency of such applications.In summary,our main contributions of this paper include:1.Reconstructing the convolutional neural network on the mobile phone by pre-trained weights.Using the mobile GPU acceleration by the OpenCL framework to develop a CNN inference library.2.Using the pruning-retraining loop to compress CNN models.Introducing Sparse Matrix-Vector Multiplication(SpMV)into the CNN inference library to enable the runtime library to support the compressed sparse CNN model.3.To take full advantage of the heterogeneous computing environment provided by current and future mobile device SoCs,this paper proposes an optimization strategy which can use heterogeneous device processors to execute the CNN in-ference.Based on the energy efficiency difference among heterogeneous proces-sors equipped on the target mobile platform,this strategy can adaptively find an energy efficient device processor combination to execute the CNN inference in parallel.4.This paper analyzes the runtime load of the CNN-based life-logging application in detail and further proposes the method of using Dynamic Voltage and Frequency Scaling(DVFS)to improve the CNN-based application performance or energy efficiency in system level.The research significance of this paper is described as follows:1.Designing and implementing a mobile CNN inference library that integrates func-tions such as model compression and heterogeneous computing task assignment.2.Proposing a strategy for parallel execution of CNN inference based on hetero-geneous device processors.This strategy can automatically evaluate the energy efficiency of heterogeneous processors on a target platform at runtime.3.Exploring strategies of using DVFS to improve the CNN-based smart application energy efficiency in system level.