ABSTRACT
One of the more lamentable results of the information-processing revolution within psychology over the past twenty years has been the replacement of the term learning by the term memory. Whereas it is sometimes difficult to distinguish the learning experiments of twenty years ago from today's memory experiments, it is increasingly clear that remembering is only one kind of learning. As long as our theories of knowledge representation were simple, this substitution caused no problem. If knowledge is essentially declarative and unstructured, new learning can be carried out by simply adding new facts to the data base. Over the past several years, however, we have been led to a significantly more complex representational theory. In particular, we have come to see knowledge as embedded in schemata that we see as largely composed to specialized bits of procedural knowledge (Bobrow & Norman, 1975; Rumelhart, in press; Rumelhart & Ortony, 1977). In a recent paper (Rumelhart & Norman, 1978), we began a logical analysis of what learning must amount to in the context of a schema-based representational system. According to our analysis, the adoption of the schema as the basic unit of knowledge representation has implicit in it three qualitatively different kinds of learning.
Accretion—the encoding of new information in terms of existing schemata. In our view, new information is interpreted in terms of relevant preexisting schemata, and some trace of this interpretation process remains after the processing is complete. This trace can serve as the basis for a later reconstruction of the original input. Thus, processing information changes the system, giving it the ability to answer questions it could not have previously answered. The system has thereby learned something new. This is presumably the most common and least profound sort of learning. Note that no new schemata are involved in this sort of learning. An organism that learned only in this way could never gain any new schemata; all learning would be in terms of instantiations of already existing schemata.
Tuning or schema evolution—the slow modification and refinement of a schema as a function of the application of the schema. Schema evolution is presumably a central mechanism in the development of expertise. With experience, an existing schema can be slowly modified to conform better and better to the sorts of situations to which it is to apply.
Restructuring or schema creation—the process whereby new schemata are created. This kind of learning, which we have called restructuring, or more recently simply structuring, involves the creation of new schemata that, through tuning, can themselves become highly refined and distinct concepts.
