机器学习算法可近似性的量化评估分析
发布时间：2019-06-22 17:09
【摘要】：近年来,以神经网络为代表的机器学习算法发展迅速并被广泛应用在图像识别、数据搜索乃至金融趋势分析等领域.而随着问题规模的扩大和数据维度的增长,算法能耗问题日益突出,由于机器学习算法自身拥有的近似特性,近似计算这种牺牲结果的少量精确度降低能耗的技术,被许多研究者用来解决学习算法的能耗问题.我们发现,目前的工作大多专注于利用特定算法的近似特性而忽视了不同算法近似特性的差别对能耗优化带来的影响,而为了分类任务使用近似计算时能够做出能耗最优的选择,了解算法"可近似性"上的差异对近似计算优化能耗至关重要.因此,选取了支持向量机(SVM)、随机森林(RF)和神经网络(NN)3类常用的监督型机器学习算法,评估了针对不同类型能耗时不同算法的可近似性,并建立了存储污染敏感度、访存污染敏感度和能耗差异度等指标来表征算法可近似性的差距,评估得到的结论将有助于机器学习算法在使用近似计算技术时达到最优化能耗的目的.
[Abstract]:In recent years, machine learning algorithms, represented by neural networks, have developed rapidly and have been widely used in image recognition, data search and even financial trend analysis. With the expansion of the scale of the problem and the increase of the data dimension, the problem of energy consumption of the algorithm is becoming more and more prominent. Because of the approximate characteristics of the machine learning algorithm itself, the technology of reducing energy consumption by approximate calculation of a small amount of accuracy of the sacrifice results has been used by many researchers to solve the problem of energy consumption of the learning algorithm. We find that most of the current work focuses on using the approximate characteristics of specific algorithms and neglects the influence of the difference of approximate characteristics of different algorithms on energy consumption optimization. In order to make the optimal choice of energy consumption when using approximate calculation of classification tasks, it is very important to understand the difference of algorithm "approximability" for approximate calculation of optimal energy consumption. Therefore, three kinds of supervised machine learning algorithms, support vector machine (SVM), random forest (RF) and neural network (NN), are selected to evaluate the approximability of different algorithms for different types of energy consumption, and some indexes, such as storage pollution sensitivity, visiting pollution sensitivity and energy consumption difference, are established to characterize the similarity gap of the algorithm. The conclusion of the evaluation will help the machine learning algorithm to optimize the energy consumption when using approximate computing technology.
【作者单位】： 计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学;
【基金】：国家自然科学基金项目(61572470,61532017,61522406,61432017,61376043,61521092) 中国科学院青年创新促进会项目(404441000)~~
【分类号】：TP181
，
本文编号：2504799
[Abstract]:In recent years, machine learning algorithms, represented by neural networks, have developed rapidly and have been widely used in image recognition, data search and even financial trend analysis. With the expansion of the scale of the problem and the increase of the data dimension, the problem of energy consumption of the algorithm is becoming more and more prominent. Because of the approximate characteristics of the machine learning algorithm itself, the technology of reducing energy consumption by approximate calculation of a small amount of accuracy of the sacrifice results has been used by many researchers to solve the problem of energy consumption of the learning algorithm. We find that most of the current work focuses on using the approximate characteristics of specific algorithms and neglects the influence of the difference of approximate characteristics of different algorithms on energy consumption optimization. In order to make the optimal choice of energy consumption when using approximate calculation of classification tasks, it is very important to understand the difference of algorithm "approximability" for approximate calculation of optimal energy consumption. Therefore, three kinds of supervised machine learning algorithms, support vector machine (SVM), random forest (RF) and neural network (NN), are selected to evaluate the approximability of different algorithms for different types of energy consumption, and some indexes, such as storage pollution sensitivity, visiting pollution sensitivity and energy consumption difference, are established to characterize the similarity gap of the algorithm. The conclusion of the evaluation will help the machine learning algorithm to optimize the energy consumption when using approximate computing technology.
【作者单位】： 计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学;
【基金】：国家自然科学基金项目(61572470,61532017,61522406,61432017,61376043,61521092) 中国科学院青年创新促进会项目(404441000)~~
【分类号】：TP181
，
本文编号：2504799
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2504799.html
