ABSTRACT: Backpropagation (Rumelhart et al., 1986a) was proposed as a general learning algorithm for multi-layer perceptrons. This article demonstrates that a standard version of backprop fails to attend selectively to input dimensions in the same way as humans, suffers catastrophic forgetting of previously learned associations when novel exemplars are trained, and can be overly sensitive to linear category boundaries. Another connectionist model, ALCOVE (Kruschke 1992), does not suffer those failures. Previous researchers identified these problems; the present article reports quantitative fits of the models to new human learning data. ALCOVE can be functionally approximated by a network that uses linear-sigmoid hidden nodes, like standard backprop. It is argued that models of human category learning should incorporate quasi-local representations and dimensional attention learning, as well as error-driven learning, to address simultaneously all three phenomena.