第一章 2 概念学习.pptVIP

2003.12.18 机器学习-概念学习 作者：Mitchell 译者：曾华军等 讲者：陶晓鹏 主要内容 概念学习简介 概念学习问题的定义 表示假设的形式 术语定义 归纳学习假设 作为搜索的概念学习 假设的一般到特殊序 Find-S：寻找极大特殊假设 Find-S算法实例1 Find-S算法实例2 Comments on Find-S 变型空间和候选消除算法 变型空间（Version Space）的定义 变型空间表示 候选消除学习算法 候选消除学习算法实例1 候选消除学习算法实例2 VS Traces VS Traces Convergence Property Computational Complexity Query Generation Using Partially Learned Concepts Learning for Multiple Categories Inductive Bias Unbiased Learning Futility of Bias-Free Learning 归纳偏置 归纳偏置（2） 归纳偏置（3） 归纳偏置（4） 小结 补充读物 Knowing the version space is useful for selecting good examples if the system is able to query an oracle for the classification of specific examples. Want an example that matches half the hypotheses in the version space so one of these halfs is eliminated if the example is positive and the other half if it is negative. After big, red circle, 1, small, red, triangle, 0 version-space is: big, red, circle, big, red, ?, big, ?, circle, ?, red, circle, ?, ?, circle, big, ?, ? example big, blue, circle matches half so is a good example to query. Given the ceiling of log2|VS|, such examples will result in convergence, if such a sequence is possible. In any case, this is the minimum number of examples needed. If the version space has not converged, then how is a novel test instance classified? If all elements of S match the instance then the entire version space must (since each element is more general than some element in S) and it can be classified as positive as confidently as in the converged case. If no element in G matches the instance then the entire version space must not match (since each element is more specific than some element in G) and it can be classified as negative as confidently as in the converged case. Otherwise, can take a vote of the hypotheses in the version space (or more tractably just the elements of S and G) to give a classification with an associated confidence value. Voting the entire version spa