Bridging levels: Using a cognitive model to connect brain and behavior in category learning

摘要: Bridging Levels: Using a Cognitive Model to Connect Brain and Behavior in Category Learning Todd M. Gureckis (tgurecki@indiana.edu) Department of Psychology, Indiana University 1101 E. 10th Street, Bloomington, IN 47405 USA Bradley C. Love (love@psy.utexas.edu) Consortium for Cognition Computation, Texas at Austin Austin, TX 78712 Abstract predict how degraded function along this circuit af- fects category learning performance these (and other) groups. In particular, SUSTAIN relates the degree preserved readily members group can individuate events, as opposed collapsing experi- ences together into common gestalt (see Figure 1). After introducing model, we explain close cor- respondence between aspects model cur- rently understood involving PFC, hippocampus, perirhinal cortex. We then review number simulations which support our the- ory. doing so, provide novel framework under- standing role plays ability. addition, analysis suggests recasting several dichotomies popular field, such distinction categorization recognition, rec- ollective familiarity-driven responding, episodic semantic memory. Mental localization efforts tend stress where more than what. argue that proper targets lo- calization are well-specified cognitive models. make case by relating an existing cat- egory hip- pocampus, perirhinal, prefrontal Results from groups varying (e.g., infants, amnesics, older adults) successfully simu- lated reducing model’s ability form new clus- ters response surprising error supervised or unfamiliar stimulus un- learning. Reported task dissociations vs. recognition) explained terms cluster recruitment demands. A major goal psychology has been de- velop understanding behavior compu- tational principles. However, often left with question what models tell us about brain. The answer is certainly not clear. growing area neuroscience offers endless source embers debate, func- tion localized described specific brain processes. focusing on mental (i.e., processes X brain?), run risk amassing list areas associ- ated certain tasks absence useful linking theories reflecting those regions interact control daily lives. paper, well-specified, process functions appropriate localization. Successful offer num- ber advantages over folk psychological, ad hoc, tra- ditional psychological theories. For example, veloped predictions, have mechanisms dynamics be related measures, simple clear starting point developing function. To conjecture, focus hu- man cortext, cortex (PFC). consider, Supervised Unsupervised STratified Adaptive Incremental Network (SUSTAIN), applied human data populations (infants, who differ their Armed its computational principles proposed mapping, able Proposed Mapping begin theory (SUSTAIN) bridge functional components Due lim- ited space, readers interested mathematical tails directed elsewhere (Love, Medin, & Gureckis, 2004). basis representing knowledge rule-based, exemplar-based, prototype-based. proposes clusters, display characteristics all three afore- mentioned approaches, underlie represen- tations. bundle features captures conjunctive relationships across dimensions. exam- ple, capture fact having wings, flying, feathers co-occur. SUSTAIN, categories represented one clusters. birds might multiple clusters natu- ral patterns regularity within song birds, prey, penguins, ostriches each separate belong super-ordinate). also mul- tiple once because they linked egories association weights adjusted during