基于对数运算的Sigmoid函数的研究与实现

Research and Implementation Based on Transcendental Function Coprocessor Sigmoid Function

基于对数运算的sigmoid函数常被用作神经网络的阈值函数,将变量映射到(0,1)之间.为了研究其高速率,高精度,低功耗的性能,以对数运算为实现原理,对sigmoid函数做了研究并继而实现其函数,通过SystemVerilog仿真平台对设计进行简单验证.通过比较仿真平台产生的输出与设计产生的输出来总结本次设计的sigmoid函数的速率,精度,功耗等性能指标.结果表明:本次设计的sigmoid函数在速率上得到了改善,进行一次函数运算需七个时钟周期(主频200 MHz)在精度上也稍有提升,降低了硬件功耗.

Sigmoid function based on the logarithm operation is often used as the threshold function of neural networkmapping variables between zero and one. In order to study its high speed, high precision, low power consumption,by logarithm operation for the realization of the principle, we do the research and then realize its function for thesigmoid function . We also do simple design verification through the SystemVerilog simulation platform, andsummarize performance indicators of the sigrnoid function on rate, precision, power consumption and so on bycomparing the output of the simulation platform with the output of design. Results show that the designs of thesigmoid function on rate are improved, a functional operation needing seven clock cycles （frequency 200 MHZ）, aslight increase on the precision. Hardware power consumption has been reduced.