关键词: 类别学习, 特征推理, 单类说, 理性模型, 贝叶斯推理

Abstract: This paper studies how feature prediction is influenced by two types of category learning at uncertain classifying circumstance. One type of learning is stimulus by stimulus and another is category by category. Anderson (1991) provided a Bayesian analysis on feather predicting when categories are uncertain. For each object containing features F and each category k, one can predict the presence of a novel feature j by using the formula: P(j\F) =Σk P(k\F)·P(j\k). That is, one calculates for the object how likely it is to be in each category k and how likely that category is to contain the property in question. Then one sums across all the categories in order to make the prediction. In short, this proposal is that people use multiple categories to make predictions when the categorization is uncertain. Murphy & Ross (1994) argued that people make category-based inductions basing on only one category, even when they are not certain that the object is in that category. They found that even if participants give a fairly low rating of their confidence in the category it does not lead them to use multiple categories at making prediction. That is, one can predict the presence of a novel feature j by using the formula: P(j\F) =P(k\F)·P(j\k).
The experiments used the learning-transfer-paradigm which has three phases: learning phase, filler phase and transfer phase. 244 participants took part in two experiments. In experiment 1, participants stopped learning until they completed 4 blocks (64 trials), and in experiment 2 until they reached an accuracy of combination of 80%. In learning phase there were two learning ways: one was stimulus by stimulus (experiment 1b and 2b) and another was categories by categories (experiment 1a and 1b) and participants reacted basing on conditions and received feedback from tester. In transfer phase participants conducted same task as learning phase except that no feedback was given during transfer.
The results in experiment 1a and 2a showed that in category by category learning way the neutral condition was rated 78% and 71.4%, comparing to 83.3% and 76.9% for the adding condition. The difference between two conditions was not significant and “t” equal to -0.43 and 0.424, p>0.05. The results in experiment 1b and 2b showed that in stimulus by stimulus learning way the neutral condition was rated 77.4% and 78.8%, comparing to 68.8% and 81.6% for the adding condition. The difference between these conditions was significant and “t” equal to 2.21 and 1.77, p <0.05.
The study demonstrated that two ways of category learning led to different category representations: in way of learning category by category participants tended to focus on information of single category and their subsequent prediction conformed to the formula: P(j\F) = P(k\F)·P(j\k). Whereas in way of learning stimulus by stimulus participants tended to focus on information of multiple categories, and their subsequent prediction conformed to the formula: P(j\F) =Σk P(k\F)·P(j\k). This observation is part of general trend that is concerned with how category learning influences category representation.

Key words: category learning, feature predicting, single category, multiple categories, Bayesian rule