具有二维状态转移结构的随机逻辑及其在神经网络中的应用

Stochastic Logics with Two-dimensional State Transfer Structure and Its Application in the Artificial Neural Network

随机计算是一种特殊的基于概率数据码流的数学计算方法,其优点在于可以采用非常简单的数字逻辑完成复杂数学运算,从而大幅降低硬件实现成本。该文首先讨论了随机计算的基本原理和主要运算逻辑,论述了传统线性状态机的不足,并分析了一种2维状态转移拓扑结构,推导了通过2维有限状态机实现高斯函数的方法。在此基础上,提出一种随机径向基函数神经网络模型,其硬件实现成本非常低,而性能与传统神经网络相当。两类模式识别实验结果显示,所提出的随机径向基函数神经网络的输出值均方误差与相应结构传统神经网络的差别小于1.3%。FPGA实验结果显示,数据宽度为12位时,随机中间神经元的电路面积仅为传统插值查表结构的1.2%、坐标旋转数字计算方法(CORDIC)的2%。通过改变输入码流长度,该神经网络可以在处理速度、功耗和准确性之间作出平衡,具有应用灵活性,适用于对成本、功耗要求较高的应用如嵌入式、便携式、穿戴式设备。

Stochastic computing is a special algorithm that performs mathematical operations with probabilistic values of bit streams rather than traditional deterministic values. The main advantage of stochastic computing is its great simplicity of hardware arithmetic units for mathematical operations to reduce the circuit cost. This paper discusses the principle of the stochastic computing and its main arithmetic logic. It analyzes a two-dimension state transition topology structure, and discusses the Gaussian function implementation method based on the two-dimension Finite State Machin （FSM）. Then, a low cost stochastic radial basis function neural network model is proposed. Results from two pattern recognition tests show that the difference of the mean squared error between the stochastic network output value and the corresponding deterministic network output value can be less than 1.3%. FPGA implementation results show that the hardware resource requirement of the proposed stochastic hidden neuron is only 1.2% of the corresponding deterministic hidden neuron with the interpolated look-up table, and is 2.0% of the CORDIC algorithm. The accuracy, speed and power of the stochastic network can be tradeoff dynamically. This network is suitable for the low cost and low power applications like embedded, portable and wearable devices.