随机门设计及其体系结构在机器学习中的应用

项目名称： 随机门设计及其体系结构在机器学习中的应用

项目编号： No.61472243

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 自动化技术、计算机技术

项目作者： 苑波

作者单位： 上海交通大学

项目金额： 80万元

中文摘要： 许多机器学习和推理问题的一个特点是存在着不确定性。在这种不确定的环境下对推理问题的解决依赖于将它转换为概率性的模型,并求解概率分布。为了降低计算的复杂度,机器学习和推理问题的解决现在多依赖于基于图模型的采样算法。进一步提高这一算法的运行性能将提升许多学习和推理问题的求解速度,增大计算机能够处理的问题的规模。然而,利用传统处理器来实现这种采样算法具有局限性。主要原因在于:1)图模型采样算法的随机性和硬件的确定性之间存在着一定的矛盾;2)基于传统处理器的实现不利于发挥算法所具备的并行性。在本项目中,我们将探索一种更高效的硬件平台来实现基于图模型的采样算法。针对算法需要以任意概率分布产生随机输出的特点,我们提出设计一种称为随机门的数字电路来更自然地产生随机输出。进一步,我们将探索如何设计基于随机门的可重构的体系结构来实现这一算法,以充分利用算法具有的并行性,并能够对任意的图进行采样.

中文关键词： 随机门；随机计算；并行计算；图模型；蒙特卡罗采样

英文摘要： Many machine learning and inference problems are characterized by uncertainties. Therefore, it is natural to model them as probabilistic problems. With this modeling, solving the inference problem is equivalent to obtaining some probability distributions. In order to reduce the computational complexity, the method of sampling a probabilistic graphical model is widely used. Accelerating this algorithm will speed up the solving of many learning and inference problems and enable problems with larger scale to be solved. However, the method of sampling a graphical model is usually implemented using a general purpose processor, which has two major drawbacks: 1) the random nature of the algorithm is incompatible with the deterministic hardware used to build the general purpose processor, and 2) the general purpose processor cannot fully exploit the inherent parallelism of the graph-based sampling algorithm. In this project, we will design a novel computational platform to efficiently implement the sampling algorithm. We propose to develop a novel type of stochastic gate to naturally generate the random samples required by the algorithm. Furthermore, we will explore how to build a reconfigurable architecture with the stochastic gates to fully exploit the parallelism of the algorithm and to make the hardware be applicable to arbitrary graphs.

英文关键词： Stchastic Gate;Stochastic Computing;Parallel Computing;Graphical Model;Monte Carlo Sampling