A category learning experiment involving human participants compared the difficulties of four types of shift learning. Initial learning was of an exclusive-or (XOR) structure on two of three stimulus dimensions. One shift type was a reversal, a second shift was to a single previously relevant dimension, a third shift was to a single previously irrelevant dimension, and a fourth shift was to an XOR on one previously relevant dimension and one previously irrelevant dimension. Results showed that reversal shift was easiest, followed, in order, by shift to a single previously relevant dimension, shift to a single previously irrelevant dimension, and a shift to a new XOR. An extended version of the ALCOVE model, called AMBRY, qualitatively fits the data. The model incorporates two essential principles. First, internal category representations that can be quickly remapped to overt responses are important for accounting for the ease of reversal shift. Second, perseverating dimensional attention is important for accounting for the ease of shifting to a previously relevant dimension as opposed to a previously irrelevant dimension. It is suggested that any model of these effects will need to implement both of these principles.

The gain of a mode in a connectionist network is a multiplicative constant that amplifies or attenuates the net input to the node. The benefits of adaptive gains in back-propagation networks are explored. It is shown that gradient descent with respect to gain greatly increases learning speed by amplifying those directions in weight space that are successfully chosen by gradient descent on weights. Adaptive gains also allow normalization of weight vectors without loss of computational capacity, and the authors suggest a simple modification of the learning rule that automatically achieves weight normalization. A method for creating small hidden layers by making hidden node gains compete according to similarities between nodes in an effect to improve generalization performance is described. Simulations show that this competition method is more effective than the special case of gain decay.>