如何增加人工神经元网络的透明度.pdfVIP

第 20 卷第 1 期 2007 年 2 月 模式识别与人工智能 PR . AI 如何增加入工神经元网络的透明度γ 胡包锅 王炼 扬双红 (中国科学院自动化研究所北京 100080) 〈中国科学院研究生院北京 100080) 曲寒冻 VoL 20 No.l Feb 2007 摘 要 针对人工神经元网络应用中最主要的商题之一一一黑箱特性进行文献综述.增加入工神经元网络系统 的透明度是解决该问题必不可少的手段.为了便于理解各种已有方法的应用特点及其局限性，提出透明度研究 中的方法分类框架.首先将透明度研究划分为两种基本策略:1)将先验信息引入系统设计; 2)从模型中提取系统 格关规则或知识.在此基础上，对各种主要方法进一步分类并进行应用特点介绍，最后对机器学习多目标研究进行 讨论.提出基于性能价格比与基于提高系统透明度的目标函数.指出提高透明度是神经元网络研究中最为 基本而又直接的解决方案.为此提出反馈知识增长(Knowledge Increasing via Feedback)型机器学习方法. 关键读 机器学习，人工神经元网络，先验知识，臼纳，演绎，黑箱 中画法分类号 TP183; TP181 How to Add Transparency to Artificial Neural Networks HU Bao-Gang , W ANG Yong , Y ANG Shuang-Hong , QU Han-Bing CInstitute of Automation , Chinese Academy of Sciences , Beijing 100080) (Graduate School , Chinese Academy of Sciences , Beijing 100080) ABSTRACT The 1血Z丑1羽lal♂in issue about Adding transparency i总swε11 recognized to be an effec盯ti忖Vεsolution tωo deali扫ng with t出hi必s probl沁εm. Significant benεf丑its are obtained tl七1rough王 USH丑19 this 丘pproach ， such as providíng a certaín degree of comprehensive power , decreasing model size , speeding learníng process and improving generalization capability. A hierarchical classification is applied to the existing approaches for better understanding of their intrinsic features and limitations. Thεfirst level of classification is made by two strategies: building prior knowledge into neural networks; extracting rules embedded within networks. Most of important approaches arεintroduced and compared in detail with further classifications within each strategy. Finally，由e pεrsonal perspectives to the studies of machine learning are presented. Other objective functions are suggested for the extension of studies , such as performance- to-cost ratio and transparency. The study of increasing transparency to ANN s is considered as the most fundamental and dirεct solution to the other existing issues. A new machine learning approach called KnowledgεIncreasing via Feedback is proposed. 美因家自然科学基金项目 (No 、国家自然