【摘要】：卷積神經(jīng)網(wǎng)絡(luò)因深度學(xué)習(xí)概念的提出再一次被研究人員所重視.激勵(lì)函數(shù)是卷積神經(jīng)網(wǎng)絡(luò)的一個(gè)重要組成部分,選取了sigmoid函數(shù)作為實(shí)驗(yàn)對(duì)象.討論了當(dāng)前幾種可行的逼近方法,最終采用分段四階多項(xiàng)式擬合sigmoid函數(shù).在FPGA上使用Verilog硬件描述語(yǔ)言設(shè)計(jì)了并行電路,并采集了數(shù)據(jù)集進(jìn)行FPGA與CPU版本caffe庫(kù)進(jìn)行運(yùn)算效率對(duì)比.實(shí)驗(yàn)結(jié)果表明,此種方法誤差小效率高,FPGA在深度學(xué)習(xí)領(lǐng)域有著廣闊的應(yīng)用前景.
[Abstract]:Convolutional neural networks have once again been paid attention to by researchers because of the concept of deep learning. The excitation function is an important part of the convolution neural network. The sigmoid function is selected as the experimental object. Several feasible approximation methods are discussed. Finally, the piecewise fourth order polynomial is used to fit the sigmoid function. The parallel circuit is designed by using Verilog hardware description language on FPGA, and the data sets are collected and compared with FPGA and CPU version caffe library. The experimental results show that this method has a wide application prospect in the field of deep learning with small error and high efficiency.
【作者單位】： 四川大學(xué)計(jì)算機(jī)學(xué)院視覺(jué)合成圖形圖像技術(shù)國(guó)家重點(diǎn)學(xué)科實(shí)驗(yàn)室;
【基金】：國(guó)家“八六三”計(jì)劃項(xiàng)目(2015AA016405) 四川省科技廳科技支撐項(xiàng)目(2016GZ0097)
【分類號(hào)】：TN791;TP183

【相似文獻(xiàn)】

相關(guān)期刊論文 前4條

1 吳訓(xùn)威,陳豪;基于觸發(fā)行為的J、K激勵(lì)函數(shù)的最小化技術(shù)[J];浙江大學(xué)學(xué)報(bào)(理學(xué)版);2004年02期

2 周定勇;何金保;陳永杰;;基于J、K激勵(lì)函數(shù)最小化方法及電路的設(shè)計(jì)[J];微型機(jī)與應(yīng)用;2013年05期

3 羅慧;王友仁;林華;姜媛媛;;任意周期激勵(lì)函數(shù)的模擬電路測(cè)試激勵(lì)優(yōu)化設(shè)計(jì)[J];電子學(xué)報(bào);2011年08期

4 嚴(yán)單貴;張承暢;;一種改進(jìn)的獲得J-K觸發(fā)器激勵(lì)函數(shù)的方法[J];微電子學(xué)與計(jì)算機(jī);2013年07期

，

本文編號(hào)：2226181