[PS-2.1] Heterarchy in computational models of cognition

Shillcock, R.

University of Edinburgh

Conventionally, cognitive scientists derive the universal categories represented in cognitive models - such as feature, phoneme or word - from formal (or informal) theorizing about the domain. These categories are then placed in hierarchies, typically following a data-path from input to some more centrally mediated output. For instance, in TRACE (McClelland & Elman, 1986) an idealized speech input leads to word identification. Churchland, Ramachandran and Sejnowski (1994) argue against the artificial informational encapsulation of particular processing domains and call for a ?heterarchical, interactive ... theory?. I discuss a means of achieving this goal by employing an alternative understanding of ?universal?. A universal is classically defined as something that unifies a number of different things. It is conventionally derived by positing something that is similar or common among a number of different specific instances. Such universals are abstract universals: they are useful for gaining initial traction on a domain, but are ultimately constrained to telling us only about ordered relations within the domain. They can be related to each other by mathematical functions and made to simulate human behaviour, but they are fated always to encounter data that they cannot accommodate. However, they can be complemented by a different type of universal, a concrete universal, a concept with a venerable philosophical history. A concrete universal is a universal by virtue of the fact that it mediates every other entity in the domain; it unifies a number of different things but is itself a real, theory-independent entity that can itself be manipulated and observed. The concrete universal is the most principled approach to defining a heterarchical processing architecture and represents a radically different approach to simplicity, explanation, and completeness. I will illustrate this position with a discussion of our current work in the modelling of speech processing in the TRACE tradition.