Competitive Learning in Biological and Artificial Neural Computation

tigate competition among the participating functional units. One kind of behavioral evidence in support of the existence of tuned units is provided by the specificity of improvement in typical learning situations to the given task and stimulus. The specificity of learning is pervasive: in vision, it occurs in the entire range of tasks from intensity grating discrimination [4] to face recognition [5]. In some cases, stimulus-specific learning has been shown to lead to a degradation of performance on a different range of stimuli in the same task. For example, the lowering of vertical vernier detection threshold with practice transfers neither to horizontal verniers, July, 1997 Competitive Learning in Biological and Artificial Neural Computation 2 nor to other acuity-related tasks [6]; moreover, subjects trained on verniers of one orientation exhibit initially a threshold that is elevated relative to the baseline when tested on verniers of the orthogonal orienta