The Emergence of Multiple Learning Systems

摘要: The Emergence of Multiple Learning Systems Bradley C. Love (love@psy.utexas.edu) Matt Jones (mattj@psy.utexas.edu) Consortium for Cognition and Computation, University Texas at Austin Austin, TX 78712 USA Abstract cumstances. For example, ATRIUM (Erickson & Kr- uschke, 1998) contains a rule exemplar learning sys- tem. Which system is operable determined by gat- ing system, allowing different classification procedures to be applied parts the stimulus space. familiar items could classified exem- plar whereas rules unfamiliar items. power apply qualitatively pro- cedures stimuli hallmark multiple systems models. Proposing begs questions how many are present do they interact. Are there two, three, or thirty-four systems? Do some tems combine outputs others shunt each other? These not trivial answer. two model may suffice one data set, but new manipulation provide evidence third system. As propagate, complexity overall dramatically increases. Building in this degree complicates evaluation. Instead proposing complex category containing systems, we advocate approach model- which emerge from flexible adaptive clustering mechanism’s interactions with environment. We evaluate hypothesis that relatively small set principles can effectively “grow” knowledge structures satisfy needs models intended address. hold separate often organized around discrepant principles, their support human cat- egorization. Rather than propose model, adopt systems’ viewpoint CLUSTer Error Reduc- tion (CLUSTER), retains flexibility characteristic building as needed learner’s goals. Importantly, CLUS- TER ostensibly dif- ferent space, mul- tiple describe simulation study CLUSTER develops cluster representations item types. Rule-following captured clusters broadly tuned focused on rule-relevant as- pects, exceptions (especially those violate high-frequency rules) narrowly focus item-specific qualities. end considering relation between findings cognitive neuroscience learning. Introduction Proposals representation diverse, ranging exemplar- (Medin Schaffer, 1978) prototype-based (Smith Minda, include proposals these extremes (Love, Medin, Gureckis, 2004). Determining best psychological difficult perform well situation bested competing situation. One possibility single “true” model. In learning, line reasoning has led development systems. more behavior reflects contributions uti- lize distinct representations. idea be- havior enjoys widespread neu- roscience (see Ashby O’Brien, 2005, review Nosofsky Zaki, 1998, dissenting opinion). detail relative component tems. depend cir- Past Work Current Challenges Previous work SUSTAIN precursor introduce here, par- tially delivered promise flexibly structures. starts simple recruits response surprising events, such encountering an unsupervised making error supervised (cf. Carpenter Grossberg, 2003). Surprising events indicative existing clus- ters current goals should grow (i.e., clusters). modified adjust position center them amidst members. Dimension-wide attention also adjusted accentuate properties most predictive across clusters.