如何增加人工神经元网络的透明度?  

How to Add Transparency to Artificial Neural Networks

作　　者：胡包钢[1,2] 王泳[1,2] 杨双红[1,2] 曲寒冰[1,2] 

机构地区：[1]中国科学院自动化研究所,北京100080 [2]中国科学院研究生院,北京100080

出　　处：《模式识别与人工智能》2007年第1期72-84,共13页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.60073007);国家自然科学基金创新集体项目(No.60121302)

摘　　要：针对人工神经元网络应用中最主要的问题之一——"黑箱"特性进行文献综述.增加人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)从模型中提取系统相关规则或知识.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.The main issue about ＂black box＂ inherent in artificial neural networks （ANN＇ s） is discussed. Adding transparency is well recognized to be an effective solution to dealing with this problem. Significant benefits are obtained through using this approach, such as providing a certain degree of comprehensive power, decreasing model size, speeding learning process and improving generalization capability. A hierarchical classification is applied to the existing approaches for better understanding of their intrinsic features and limitations. The first level of classification is made by two strategies： building prior knowledge into neural networks; extracting rules embedded within networks. Most of important approaches are introduced and compared in detail with further classifications within each strategy. Finally, the personal perspectives to the studies of machine learning are presented. Other objective functions are suggested for the extension of studies, such as performanceto-cost ratio and transparency. The study of increasing transparency to ANN ＇s is considered as the most fundamental and direct solution to the other existing issues . A new machine learning approach called Knowledge Increasing via Feedback is proposed.

关 键 词：机器学习 人工神经元网络 先验知识 归纳 演绎 黑箱 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP181[自动化与计算机技术—控制科学与工程]