Abstract
To optimally exploit the statistical structure in signals produced by the visual environment, the visual system ought to learn when a signal that was previously uninformative becomes useful for inferring a property of the world. If the visual system can learn to use a newly reliable signal, one would expect to see effects of the new signal on both decisions at threshold (Ernst & Jäkel, VSS 2003) and on visual percepts. We designed a test of these propositions, using a task in which observers judge direction of motion-in-depth for a stereoscopically defined surface. The signals for motion-in-depth (vergence and looming) can be precisely controlled. The surface rotates about the cyclopean line of sight, at an angular velocity that is either proportional to speed in depth, or else uncorrelated with speed in depth, depending on the training group to which the observer is assigned; this rotation is the new cue for motion-in-depth. On a small number of probe trials in the Correlated condition, angular velocity is reversed. If rotation is recruited as a new cue by the visual system, it will influence both thresholds and motion-in-depth judgments. Data from correlated and uncorrelated training conditions will be reported. We make several points: (1) Adaptation phenomena (“normalizations”) that recalibrate perception relative to a previously learned Bayesian prior are different from cue recruitment, and cause negative, rather than positive, adaptation aftereffects. (2) We suppose that the visual system can learn that signals are uncorrelated. An ideal learner ought therefore to learn new associations between signals slowly, if at all, if the signals have been reliably uncorrelated in the past. (3) Learning may be facilitated by practice at threshold, where the new signal is actually of use to the visual system in making discriminations. (4) A complete theory would need to explain the extent to which learning is context dependent.
EY013988