本文選題：分段非線性逼近法　切入點(diǎn)：Sigmoid函數(shù)　出處：《電子技術(shù)應(yīng)用》2017年08期 　論文類型：期刊論文

【摘要】：使用分段非線性逼近算法計(jì)算超越函數(shù),以神經(jīng)網(wǎng)絡(luò)中應(yīng)用最為廣泛的Sigmoid函數(shù)為例,結(jié)合函數(shù)自身對(duì)稱的性質(zhì)及其導(dǎo)數(shù)不均勻的特點(diǎn)提出合理的分段方法,給出分段方式同逼近多項(xiàng)式階數(shù)對(duì)逼近結(jié)果精度的影響。完成算法在FPGA上的硬件實(shí)現(xiàn),給出一種使用三階多項(xiàng)式處理Sigmoid函數(shù)的擬合結(jié)果及流水線架構(gòu),處理精度達(dá)到10-5數(shù)量級(jí),最大頻率達(dá)到127.327 MHz,滿足了高速、高精度的處理要求。
[Abstract]:The piecewise nonlinear approximation algorithm is used to calculate transcendental function. Taking Sigmoid function, which is widely used in neural network, as an example, a reasonable piecewise method is proposed according to the property of symmetry of function itself and its uneven derivative. The effect of piecewise method and order of approximation polynomial on the precision of approximation results is given. The hardware implementation of the algorithm on FPGA is given, and a fitting result of Sigmoid function using third-order polynomial and pipeline structure are given. The processing accuracy reaches 10-5 orders of magnitude. The maximum frequency is 127.327 MHz, which meets the requirements of high speed and high precision processing.
【作者單位】： 合肥工業(yè)大學(xué)微電子設(shè)計(jì)研究所;
【分類號(hào)】：O174;TN791
，

本文編號(hào)：1573467