河海大学学报(自然科学版)
河海大學學報(自然科學版)
하해대학학보(자연과학판)
JOURNAL OF HOHAI UNIVERSITY (NATURAL SCIENCES)
2013年
4期
360-364
,共5页
敏感性%神经网络%单隐层感知机
敏感性%神經網絡%單隱層感知機
민감성%신경망락%단은층감지궤
sensitivity%neural network%single hidden-layer perceptron
为使MLP神经网络对权扰动的敏感性计算算法更符合实际需求,提出一个不依赖隐层向量相互独立假设的敏感性计算算法。该算法通过自底向上依次计算单个神经元和单个层的敏感性来获得整个网络的敏感性,并采用数值计算方法给出敏感性的数学表达式。以UCI数据集为基础,验证该算法的计算精度,结果表明该算法得到的敏感性理论值能较好地逼近敏感性的模拟值。
為使MLP神經網絡對權擾動的敏感性計算算法更符閤實際需求,提齣一箇不依賴隱層嚮量相互獨立假設的敏感性計算算法。該算法通過自底嚮上依次計算單箇神經元和單箇層的敏感性來穫得整箇網絡的敏感性,併採用數值計算方法給齣敏感性的數學錶達式。以UCI數據集為基礎,驗證該算法的計算精度,結果錶明該算法得到的敏感性理論值能較好地逼近敏感性的模擬值。
위사MLP신경망락대권우동적민감성계산산법경부합실제수구,제출일개불의뢰은층향량상호독립가설적민감성계산산법。해산법통과자저향상의차계산단개신경원화단개층적민감성래획득정개망락적민감성,병채용수치계산방법급출민감성적수학표체식。이UCI수거집위기출,험증해산법적계산정도,결과표명해산법득도적민감성이론치능교호지핍근민감성적모의치。
In this study , an algorithm was developed to compute the sensitivity of a single hidden-layer perceptron to weight perturbation .This algorithm , which only requires that the first layer ’ s input elements be independent of each other but which has no such a restriction on the hidden layer , is more adaptable to real applications .The sensitivity was computed bottom-up: the sensitivity of a neuron was considered first and the entire network was considered subsequently . The sensitivity expression was obtained with the numerical calculation method . Experimental results on UCI datasets verified the precision of the proposed algorithm . The computed value approached the theoretical value .