The emergence of multiple learning systems

Abstract

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