We compare improvement through training in vernier acuity under different feedback conditions in order to clarify the role of feedback during learning of a perceptual task and to test different (neural network) models of perceptual learning. Improvement of performance is measured in 49 observers under feedback, no feedback, uncorrelated feedback, partial feedback, and block feedback conditions. Correct feedback conditions yield a larger improvement of performance than manipulated and no feedback conditions. Providing feedback that is uncorrelated to the observers' responses prevents learning, while the effect of block feedback does not differ significantly from complete feedback. Our results cannot be explained by learning rules that depend exclusively on an external teacher or by models that propose learning in an exposure-dependent way with unsupervised learning rules but without top-down influences.