In the preceding chapters, we have discussed the psychological foundations and technical realization of a model of demand-driven concept formation that exploits the problem-solving context of the reasoner to help it decide which new concepts to introduce into the representation. The method constructs new concepts out of concepts (or predicates) already existing in the representation by combining them into rules that express sufficient or necessary conditions on concept membership. This means that the approach must assume the elementary building blocks of concepts as given, and seems unable to account for the origin of truly new features or concepts that are not combinations of existing ones.
KeywordsHide Layer Connectionist Network Concept Formation Categorical Representation Hide Unit
Unable to display preview. Download preview PDF.
- 1.We have used, among others, the reviews by Gibson and Spelke [Gibson and Spelke, 1983] about perception, Mandler [Mandler, 1983] about representation, Sigel [Sigel, 1983] about concepts, and by Oerter and Montada [Oerter and Montada, 1982].Google Scholar
- 2.Indeed connectionist networks and symbolic algorithms can be applied and compared on the same tasks [Shavlik et al., 1991].Google Scholar
- 3.Incidentally, in the actual model described in [Hamad, 1990; Hamad, 1992], the idea of context is not modeled, whereas it is captured by many “symbolic” learning algorithms, like the TDIDT family [Quinlan, 1983] or recent clustering algorithms (e.g. [Lebowitz, 1987; Fisher, 1987a; Kietz and Morik, 1994]).Google Scholar
- 4.The same is true for the model of Cottrell et. al. [Cottrell et al., 1990].Google Scholar
- 5.Suggested by K. Morik.Google Scholar