Human Object Perception as Bayesian Inference

Dan Kersten
University of Minnesota

The certainty of human visual experience stands in striking contrast to
the objective ambiguity of natural images. Similar objects can produce
very different retinal images, while quite different objects can give
rise to very similar images. In ways yet to be understood, our brains
are specialized for the accurate and reliable interpretation of natural
images. The Bayesian framework provides tools, insights, and theories
for studying how the human visual system may resolve objective
ambiguity. Bayesian decision theory specifies how multiple image
features, prior knowledge about visual objects and scenes, and task
demands can be combined to infer useful interpretations of the image.
Bayesian models of image formation, "external" generative models, play
a useful role in studies of perceptual inference. They characterize
and simplify complex causal relationships in the generation of image
features, which in turn help to constrain models and algorithms for
"inverse" inference. Although there has been considerable progress
using Bayesian methods in computer vision and human psychophysics, we
know relatively little about possible cortical mechanisms through which
prior information is acquired and later combined with incoming local
image measurements. Does the visual cortex use "internal" generative
models that in some sense recapitulate external generative models in
order to infer object properties? For example, are global stored
object representations used to predict incoming local image
measurements in order to resolve ambiguities of occlusion and local
motion? I will describe some recent functional magnetic resonance
imaging studies of human cortical activity that have implications for
theories of perceptual inference in object shape perception.

Presentation Files (Zip Archive)

Back to Probabilistic Models of Cognition: The Mathematics of Mind