Relational Perception and Cognition: Implications for Cognitive Architecture and the Perceptual-cognitive Interface

A fundamental aspect of human intelligence is the ability to represent and reason about relations. Examples of relational thinking include our ability to appreciate analogies between different objects or events (Gentner, 1983; Holyoak & Thagard, 1995), our ability to apply abstract rules in novel situations (e.g., Smith, Langston & Nisbett, 1992), our ability to understand and learn language (e.g., Kim, Pinker, Prince & Prasada, 1991), our ability to learn and use categories (Ross, 1987), and even our ability to appreciate perceptual similarities (e.g., Palmer, 1978; Goldstone, Medin & Gentner, 1991; Hummel, 2000a; Hummel & Stankiewicz, 1996a). Relational inferences and generalizations are so commonplace that it is tempting to assume that the psychological mechanisms underlying them are relatively simple. But this would be a mistake. The capacity to form and manipulate explicit relational (i.e., symbolic) representations appears to be a late evolutionary development (Robin & Holyoak, 1995), closely tied to the substantial increase in the size and complexity of the frontal cortex in the brains of higher primates, most notably humans (Stuss & Benson, 1987). A review of computational models of perception and cognition also suggests that the question of how we represent and reason about relations is nontrivial (see Hummel & Holyoak, 1997, 2003): Traditional symbolic models of cognition (e.g., Anderson, Libiere, Lovett, & Reder, 1998; Anderson, 1990; Falkenhainer, Forbus & Gentner, 1989) simply assume relations as a given, making no attempt to understand the origins or detailed nature of these representations in the neural substrate; and traditional connectionist/neural networks models (e.g., Edelman & Intrator, 2003; St. John & McClelland, 1992; O’Reilly & Rudy, 2001; Riesenhuber & Poggio, 1999) fail to represent relations at all. Indeed, the proponents of such models typically reject the idea that the human cognitive apparatus is capable of representing relations explicitly (see Hummel, 2000; Hummel & Holyoak, 1997, 2003, for reviews). Comparatively few models have attempted to address the question of how a neural architecture can represent and process relational structures, or the related question of how early, non-relational representations and processes, for example in early vision, make contact with later, more explicitly relational/symbolic representations (e.g., as underlie reasoning; see Gasser & Colunga, 2001; Hummel & Biederman, 1992; Hummel & Holyoak, 1997, 2002; Shastri & Ajjenagadde, 1993; Strong & Whitehead, 1989).