When learning to categorize multidimensional stimuli, humans learn which stimulus dimensions are relevant or irrelevant for distinguishing the categories. A well-established result is that category distinctions requiring only one of two potentially relevant dimensions, called "filtration" tasks, are easier to learn than categorizations requiring both dimensions, called "condensation" tasks (Garner 1974; Posner 1964). In this article, three connectionist models of category learning are fit to newly obtained human learning curves. It is found that ALCOVE (Kruschke 1990, 1992), with its dimensional attention learning mechanism, shows an advantage for filtration, but standard back propagation (Rumelhart, Hinton & Williams 1986) and the configural-cue model (Gluck & Bower 1988a, 1988b) do not. New versions of those models, that include mechanisms for attention learning, do show an advantage for filtration. It is then shown that the extended versions suffer other problems: Extended back propagation suffers catastrophic forgetting (McCloskey & Cohen 1989; Ratcliff 1990), as shown by fits to newly obtained data, and the extended configural-cue model "attends" to dimensions ignored by humans. ALCOVE suffers neither of those problems in the present situations. Two conclusions are reached: First, connectionist models of category learning must incorporate some form of dimensional attention learning. Second, when the stimuli can be appropriately represented as points in a multi-dimensional psychological space, then a connectionist model should encode them with smooth local representations, and not with the broadly distributed representation in standard back propagation or with the discrete-valued representation in the configural-cue model.