Cascade-correlation as a model of representational redescription

How does knowledge come to be manipulable and flexible, and transferable to other tasks? These are issues which remain largely untackled in connectionist cognitive modelling. The Representational Redescription Hypothesis (RRH) (Karmiloff-Smith, 1992b) presents a framework for the emergence of abstract, higher-order knowledge, based on empirical work from developmental psychology. The RRH claims that during learning/development initiallyimplicit knowledge is rendered progressively more explicit via the reiterated action of the redescription process, resulting in a hierarchy of increasingly explicit and accessible representations. This thesis focuses on investigating in practice claims made for connectionism as a model of redescription (e.g., Clark and Karmiloff-Smith (1993)) and on applying methods from recent work in developmental connectionism to the construction of a computational model of RR. The modelling effort centres on a constructive incremental architecture — cascade-correlation (CC) (Fahlman& Lebiere, 1990)—which produces a conservative hierarchy of increasingly high-level representations as the RRH proposes. Two main models are presented. The first is designed to capture a feature of children’s comprehension of the French article system (Karmiloff-Smith, 1979a). Redescriptive effects are seen here in the changing functional status of article representations as well as in symptomatic behavioural errors. Resource-phasing was also applied to two important internal parameters of CC. The second model aims to capture the effects of redescription on sequence learning. Recurrent CCwas trained to count, and to give and compare the cardinalities of small series of stimuli. Accessibility of representations was assessed here through task transfer. Despite some success in capturing transfer, a short complementary study of structural transfer between networks learning formal grammars suggested that positive transfer in CC depends on perceptual similarity as in other supervised connectionist schemes. Themodels also address constraints on RR such as the timing and triggering of redescription and the ordering of representational formats. A brief comparative study of skeletonisation is also presented. Submitted for the degree of D. Phil. University of Sussex