Abstract
The combination of sensory cues is statistically optimal in several cases (Ernst & Banks, 2002; Alais & Burr, 2004) in the sense of reducing the variance of the resulting combined percept. Previously, these studies have assumed that sensory estimates are combined fully into a unified percept and access to single-cue estimates is lost. Here we examine a more general Bayesian view of cue combination, in which a “coupling prior” (Ernst, 2005) can probabilistically characterize both the potential for optimal improvements in perceptual accuracy as well as the possibility of partial cue combination. The variance of the coupling prior for two cues is presumably determined by their natural co-variation statistics and their degree of consistency over time. As the variance of the coupling between two cues increases, there should be greater access to single-cue information, less improvement in discriminability, and a faster rate of adaptation (see Burge, Girshick, Banks, VSS2007). We tested the first two predictions for each of two situations: disparity and texture slant cues and visual and haptic size cues. In one experiment, observers matched a two-cue conflict stimulus to a single-cue stimulus. We varied the single-cue reliability and the conflict magnitude between the cues. This allowed us to determine the extent to which the two cues were combined into a unified percept and thereby to estimate the variance of the coupling prior. In another experiment, we measured discrimination thresholds as observers matched a two-cue conflict stimulus to a two-cue, no-conflict stimulus under similar conditions. The extent of improvement in thresholds also yields a constraint on the coupling prior variance estimate. The results of both experiments suggested that the variance of the coupling prior is smaller for disparity and texture slant than for visual and haptic size.
NIH RO1 EY012851, AOF William C. Ezell Fellowship for JB