Learning by Example

Abstract

Everyone has many personal experiences of learning by example. While much psychological research has investigated “concept learning” (cf. Bruner, Goodnow, & Austin, 1956; Hayes-Roth & Hayes-Roth, in press; Hunt, 1952), that rubric is too narrow to embrace the variety of situations in which learning by example occurs. A brief list of such situations includes:

• 1. Traditional concept learning, such as inducing the class characteristics of “triangle”: “Three distinct line segments such that each line segment is coterminous, with a different line segment at each of its endpoints.” Such a rule can be induced from various examples of triangles; all examples necessarily manifest the rule, although they may differ from one another in irrelevant ways (e.g., in absolute and relative size, shape, orientation, color, texture). Note that although most traditional concept learning tasks employ only attribute-value descriptions, this learning task requires higher order, relational logic to characterize the structural constraints among the lines of a triangle.

• 2. Serial pattern learning, such as predicting the next item in a conceptually organized sequence. Traditional research on this problem has centered on mathematical sequences of symbols and various algorithmic models of memory processes for simulating the sequence generator. Other examples of this type of behavior include anticipation of expectable events (e.g., words or topics in a text that are predictable from preceding context) and prediction of cyclic phenomena. In these situation, subsequences of the preceding sequence of items serve as examples from which the sequence generation rule is induced.