Adaptive behavioural cognition.

Abstract

Cognitive Science is at a crossroad. Since its inception, the prevailing paradigm in Cognitive Science (and associated fields such as Artificial Intelligence, Cognitive Psychology, and Linguistics), has been a formal, computer-based model of cognition - often termed the Symbol Processing System model (SPS) or cognitivism. This view, while still accepted by the majority of researchers, has been dogged by persistent and cutting criticism by various authors over many years. As well, the initial over-inflated promises made by the early practitioners within these fields have not come to fruition, and the initial enthusiasm has in many cases been reduced to frustration.Many researchers have looked to the field of connectionism as a solution, and this discipline has found a new lease of life after a serious setback in the early 1970s. The major emphasis within this area has been on feed-forward neural networks (FFNN), but this paradigm also has its detractors.In this thesis we critically evaluate both of these research programs, especially that of SPS. We propose a new model of human and animal cognition, termed Adaptive Behavioral Cognition (ABC), which integrates many current views on cognition, and provides a single-architecture, biologically-feasible theory that overcomes many of the problems associated with current models. As well as being an accurate description of the processes relevant to the new model, the term ABC is a none-too-subtle reference to the fact that we need to closely re-examine the aims and achievements of Cognitive Science and return to basic empirical findings in developing a theory of cognition.The new model synthesises, unifies and links together many previously disjoint ideas and observations, from the neural level through to neurological structures and to observed behaviour. The claims that we make of the model are that it is biologically and neurologically consistent and reasonable, and that it has properties more closely associated with the actual brain than either the computational (cognitivist) approach, or the simplistic FFNN. Further, the model is internally consistent and self-similar, and is consistent with the observed neuroanatomical structures of the cortex. It also provides for massive parallelism, yet retains a serial component through its use of temporal sequences.The ABC proposal outlined in this thesis takes the view that the processes of the brain are to learn associated and temporally connected sequences, rather than 'facts' or 'representations', and that the learned behaviours resulting from the associated temporal sequences are the means of cognition, rather than computational operations on representations.