A model is presented for unsupervised learning of low level vision tasks, such as the extraction of surface depth. A key assumption is that perceptually salient visual parameters (e.g., surface depth) vary smoothly over time. This assumption is used to derive a learning rule that maximizes the long-term variance of each unit's outputs, whilst simultaneously minimizing its short-term variance. The length of the half-life associated with each of these variances is not critical to the success of the algorithm. The learning rule involves a linear combination of anti-Hebbian and Hebbian weight changes, over short and long time scales, respectively. This maximizes the information throughput with respect to low-frequency parameters implicit in the input sequence. The model is used to learn stereo disparity from temporal sequences of random-dot and gray-level stereograms containing synthetically generated subpixel disparities. The presence of temporal discontinuities in disparity does not prevent learning or generalization to previously unseen image sequences. The implications of this class of unsupervised methods for learning in perceptual systems are discussed.