Learning at Different Levels of Abstraction

摘要: Learning at Different Levels of Abstraction Bradley C. Love Department Psychology The University Texas Austin Austin, TX 78712 USA love@psy.utexas.edu Abstract Previous category learning research and the SUSTAIN (Super- vised Unsupervised STratified Adaptive Incremental Net- work) model suggest that preferred cate- gory level (in a hierarchy categories) shifts towards lower- (i.e., more specific) categories when stimuli are perceived to be distinctive. This shift is in accord with work ex- pertise. In their domain expertise, experts excel (relative novices) classifying lower levels, but advantage attenuated higher-level categories. described here directly tests (within single study) this predicted interaction between stimulus dis- tinctiveness using well controlled artificial stimuli. results consistent prior utilizing natural also informative for evaluating whether attention dimension-wide all items represented common multi-dimensional space same extent) or cluster specific different conceptual clusters can stress stimu- lus dimensions so aspects stressed). not dimension- wide. Instead, implications these findings models discussed. Introduction Humans frequently utilize acquire knowledge multiple levels abstraction. For example, ob- ject classified as vehicle, car, 1978 Lin- coln Continental. Rosch, Mervis, Gray, Johnson, Boyes- Braem (1976) argue objects most easily intermediate which effectively parti- tions world into clusters. However, Tanaka Taylor (1991) have found groups people tend prefer abstraction pre- ferring lower-level narrower finer grained compared novices who cat- egories broader abstract categories). One adult humans face perception. Medin, Dewey, Murphy (1983) faster associate unique names photographs nine female faces than they catego- rize two logical struc- ture shown Table 1. possible explanation relative ease identification used Medin et al. were rich distinct, varying along many listed 1, such shape face, type nose, etc.. id- iosyncratic information makes each item 1: structure condition. four hair color, smile type, length, shirt color. Category A B tinct. Experts may sensitive idiosyncratic informa- tion novices. absence information, wisdom holds should harder learning. other words, egory acquisition interacts nature (with being lowest abstraction) becomes easier higher become distinct. Results from literature support conclusion. Shepard, Hovland, Jenkins (1961) trained subjects on six problems Ta- ble 2 Type I was easiest master, fol- lowed by II, followed Types III-V, VI. IV problem has family resemblance resembles (1983). problem, consists an underlying prototype (111 “A” 222 “B”) any matches out three member correspond- ing category. difficult master VI while requires memorize eight because no regularities exist across pair dimensions. data, al.’s distinctiveness interact be- facilitated distinctive capture interaction. (Supervised Adap- tive Network) successfully fit Shepard data set