دانلود رایگان ترجمه مقاله یادگیری ماشین کارآمد برای داده های بزرگ (نشریه الزویر 2015)

یادگیری ماشین کارآمد برای کلان داده ها: یک مقاله مروری

Efficient Machine Learning for Big Data: A Review

Abstract

With the emerging technologies and all associated devices, it is predicted that massive amount of data will be created in the next few years – in fact, as much as 90% of current data were created in the last couple of years – a trend that will continue for the foreseeable future. Sustainable computing studies the process by which computer engineer/scientist designs computers and associated subsystems efficiently and effectively with minimal impact on the environment. However, current intelligent machine-learning systems are performance driven – the focus is on the predictive/classification accuracy, based on known properties learned from the training samples. For instance, most machine-learning-based nonparametric models are known to require high computational cost in order to find the global optima. With the learning task in a large dataset, the number of hidden nodes within the network will therefore increase significantly, which eventually leads to an exponential rise in computational complexity. This paper thus reviews the theoretical and experimental data-modeling literature, in large-scale data-intensive fields, relating to: (1) model efficiency, including computational requirements in learning, and data-intensive areas’ structure and design, and introduces (2) new algorithmic approaches with the least memory requirements and processing to minimize computational cost, while maintaining/improving its predictive/classification accuracy and stability.

1 Introduction

Today, it’s no surprise that reducing energy costs is one of the top priorities for many energy-related businesses. The global information and communications technology (ICT) industry that pumps out around 830 Mt carbon dioxide (CO2) emission accounts for approximately 2 percent of the global CO2 emissions [1]. ICT giants are constantly installing more servers so as to expand their capacity. The number of server computers in data centers has increased sixfold to 30 million in the last decade, and each server draws far more electricity than its earlier models [2]. The aggregate electricity use for servers had doubled between the years 2000 and 2005 period, most of which came from businesses installing large numbers of new servers [3]. This increase in energy consumption consequently results in higher carbon dioxide emissions, and hence causing an impact on the environment. Furthermore, most of these businesses, especially in an uncertain economic climate are placed under the pressure to reduce their energy expenditure in order to remain competitive in the market [4].

With the emerging of new technologies and all associated devices, it is predicted that there will be as much data created as was created in the entire history of planet Earth [5]. Given the unprecedented amount of data that will be produced, collected and stored in the coming years, one of the technology industry’s great challenges is how to benefit from it. During the past decade, mathematical intelligent machine-learning systems have been widely adopted in a number of massive and complex data-intensive fields such as astronomy, biology, climatology, medicine, finance and economy. However, current intelligent machine-learning-based systems are not inherently efficient or scalable enough to deal with large volume of data. For example, for many years, it is known that most non-parametric and model-free approaches require high computational cost to find the global optima. With high-dimensional data, their good data fitting capacity not only makes them more susceptible to the generalization problem but leads to an exponential rise in computational complexity. Designing more accurate machine-learning systems so as to satisfy the market needs will hence lead to a higher likelihood of energy waste due to the increased computational cost.

Nowadays, there is a greater need to develop efficient intelligent models to cope with future demands that are in line with similar energy-related initiatives. Such energy-efficient-oriented data modeling is important for a number of data-intensive areas, as they affect many related industries. Designers should focus on maximum performance and minimum energy use so as to break away from the traditional’ performance vs. energy-use’ tradeoff, and increase the number and diversity of options available for energy-efficient modeling. However, despite the fact that there is a demand for such efficient and sustainable data modeling methods for large and complex data-intensive fields, to our best knowledge, only a few of these literatures have been proposed in the field [6,7].

This paper provides a comprehensive review of state-of-the-art sustainable/energy-efficient machine-learning literatures, including theoretical, empirical and experimental studies pertaining to the various needs and recommendations. Our objective is to introduce a new perspective for engineers, scientists, and researchers in the computer science, and green ICT domain, as well as to provide its roadmap for future research endeavors.

This paper is organized as follows. Section 2 introduces the different large-scale data-intensive areas and discusses their structure and nature, including the relation between data models and their characteristics. Section 3 discusses the issues in current intelligent data modeling for sustainability and gives recommendations. Section 4 concludes the paper.