Adaline对权扰动敏感性的近似计算  

Approximate Computation of Sensitivity of Adaline to Weight Perturbation

作　　者：王炳辉[1] 曾晓勤[1] 钟水明[1] 

机构地区：[1]河海大学计算机与信息工程学院,南京210098

出　　处：《模式识别与人工智能》2009年第3期354-359,共6页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金资助项目(No.60571048;60673186)

摘　　要：主要讨论Adaline对权扰动敏感性的计算.鉴于Adaline的输入和输出的不连续性,敏感性定义为Adaline对于所有可能输入在权值发生扰动的情况下输出发生变化的概率.借助超球面模型和解析几何的技术,给出一个近似计算Adaline敏感性的方法.在输入维数足够大的情况下,该方法优于以往的其它方法.它在牺牲少量计算精度的情况下,极大地降低了计算复杂度,使得敏感性更具实用价值.The sensitivity of Adaline to weight perturbation is discussed. Considering the discrete feature of input and output of Adaline, the sensitivity is defined as the probability of an Adaline＇s output inversion due to the weight perturbation with respect to all possible inputs. By hypersphere model and analytical geometry technique, a method is proposed for approximately computing the sensitivity. Under the circumstance of high enough input dimension, the method is prior to other existing methods. It reduces computational complexity greatly with little sacrifice in precision and makes the sensitivity more practical.

关 键 词：敏感性 权扰动 超球面 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]