Abstract

A model of a visual system takes inputs from a receptor array and provides estimates of properties of a scene: collinearity, etc. Such a model is calibrated if its estimates reflect the actual state of the scene: straight lines, for example, are judged to be straight. Ahumada and Yellott raise the issue of how a visual system can be calibrated to compensate for a change in the locations of receptors in the sensing array. If a visual system is properly calibrated, eye movements can be compensated for transformations corresponding to rigid shifts of the sensing array. Changes in the state of the visual system following an eye movement that cannot be compensated for by such a transformation indicate that the visual system is miscalibrated. I show that such discrepancies are sufficient information to drive a calibration algorithm that operates without external feedback (unsupervised learning).

© 1988 Optical Society of America