ALCOVE is a connectionist model of category learning that incorporates an exemplar-based representation (Medin & Schaffer, 1978; Nosofsky, 1986) with error-driven learning (Gluck & Bower, 1988a,b; Rumelhart, Hinton & Williams, 1986). Like Nosofsky's generalized context model (GCM),ALCOVE assumes that stimuli can be represented as points in a multi-dimensional similarity space, and that each dimension has an attention strength. ALCOVE extends the GCM by adding a learning mechanism for the dimensional attention strengths and by using error-driven learning for the association weights between exemplars and categories. ALCOVE extends the network models of Gluck and Bower by allowing continuous stimulus dimensions and including explicit dimensional attention learning. Dimensional attention learning allows ALCOVE to fit human performance in situations when some stimulus dimensions are irrelevant (Shepard, Hovland & Jenkins, 1961) or when dimensions are correlated (Medin, Altom, Edelson & Freko, 1982), unlike other error-driven network models. And error-driven learning allows ALCOVE to fit human data showing apparent base-rate neglect where other exemplar-based models could not (Gluck & Bower, 1988a,b). The exemplar-based representation in ALCOVE also spares it from the catastrophic forgetting seen in standard back propagation (McCloskey & Cohen, 1989; Ratcliff, 1990), despite the fact that ALCOVE is an error-driven feed-forward network model. Finally, it is shown ALCOVE can generate three-stage learning of rule-described and exceptional cases (cf. Rumelhart & McClelland, 1986), as a natural consequence of dimensional attention learning.