The authors describe how a simple linear associative model with a novel learning rule is used to learn psychologically plausible shape descriptors of simple shapes such as characters, digits, and electronic circuit components. Their results show the power of teaching a neural network to associate general-purpose features with categories instead of discovering these features after trial and error. The use of general-purpose features and the proposed learning rule make it possible to teach the system to discriminate with an accuracy of 94% for digits, characters, and electronic gates with about eight training examples/character. The main advantage in using general-purpose features is the invariance to size, and the speed of recognition and learning. The recognition and learning take about a second on a Sun-3 workstation. Unlike J.A. Anderson and M. Mozer's (1981) model, this model does not overgeneralize, but learns to distinguish between distinct shapes that map on to the same abstract category. The relative importance of the features used in recognition is also discussed.< >