基于FPGA的具有抗衰老机制的机器学习

项目名称： 基于FPGA的具有抗衰老机制的机器学习

项目编号： No.60870001

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 汪玉

作者单位： 清华大学

项目金额： 29万元

中文摘要： 机器学习应用广泛，其计算平台的速度成为机器学习发展的主要推动力之一。同时，随着硬件平台受到可靠性机制，例如老化机制和软错误的影响，平台可靠性成为将成为机器学习领域的一个热点问题。本项目将研究基于FPGA的"抗衰老"的机器学习。首先发掘机器学习算法的并行性，进行多层次的可靠性和性能分析；其次，为各种机器学习算法建立"抗衰老"的FPGA功能单元和连接方式库，本项目将基于库的选择的方法，提出一种基于FPGA的可靠的机器学习的实现方法，高效地利用FPGA对机器学习算法进行加速；同时，本项目将根据平台运行时的特性，动态的进行重配置，进一步提高机器学习的寿命；最后，基于FPGA的"抗衰老"机器学习的共性，探索是否存在适合机器学习这个应用领域的新的体系架构。本项目的研究将可以有效地解决机器学习这个应用领域所面临的速度/可靠性/设计复杂度问题，推动机器学习研究的发展，使其更快更好为国民经济建设服务。

中文关键词： 机器学习；FPGA；抗衰老；老化机制；

英文摘要： As machine learning applications are widely used, the calculation speed of machine learning algorithms is becoming one of the main driving force. Meanwhile, with the reliability issues of the hardware platforms, such as the aging mechanisms and the impact of soft errors, the platform reliability of machine learning will become a hot issue. This project will study the FPGA-based "anti-aging" machine learning. Firstly, the parallelism will be explored and multi-level reliability and performance will be analyzed. Secondly, for a variety of machine learning algorithms, "anti-aging" FPGA functional unit and the connection library will be established. Based on the method of choosing modules from the library, an FPGA-based implementation and acceleration of reliable machine learning will be proposed. At the same time, the project will study the characteristics of run-time dynamic reconfiguration to further improve the life of machine learning. Finally, a general and efficient FPGA-based "anti-aging" machine learning framework will be explored. The research project will be able to effectively address the speed / reliability / design complexity issues of the machine learning applications and promote the development of machine learning research, to make it faster and better services for the national economy.

英文关键词： Machine learning;FPGA;anti-degradation;aging;