Schema-based learning

Abstract

Schema-based learning (SBL) builds on the schema theory of The Metaphorical Brain 2 (TMB2) in a number of ways. In SBL we incorporate not only schemas which provide units of interaction of the organism with its environment, but also predictive schemas reflecting expectations about those interactions. These internal models exist at all levels of granularity forming a hierarchical network of schemas. Schemas as generic specifications go beyond any specific implementation (neural networks in particular). Thus, the definition of schemas must be general enough so that it applies to a whole variety of possible implementations. There have been several attempts at formalizing schemas. Lyons and Arbib [7] formalized schemas as port automata and Corbacho and Arbib [4] extended this formalization by including activity variables for each port automaton along with the corresponding dynamics. Since schemas must be general by definition we claim the “power” comes from the operations at this schema level. For instance. SBL includes schema instantiation, schema assertion by distributed competitive and cooperative dynamics over the schemas activity variables, schema tuning and schema construction (where predictive schemas play a very important role in deciding when to construct a new schema). Arbib asserts that “schema theory attempts to bridge between structure and function at the highest level”. A schema may refer to either a set of regularities in observed behavior or the internal mechanisms which give rise to such regularities. In analyzing a living system we mostly refer to the first. In designing (synthesizing) a system we must provide the second, i.e., provide internal mechanisms. The ultimate challenge is to match both. Hanson argues that “one might posit different types of schemas, ones perhaps that vary in abstractness or perhaps have some domain or topic dependency . ..‘I. and indeed SBL introduces several types of schemas: perceptual, motor, sensorimotor, goal and predictive schemas. Current extensions on progress also include autonomous construction of incrementally abstract schemas grounded on more primitive sensorimotor schemas, much in the way Piaget [9] described the appearance of more abstract schemas along the