Abstract
Perceptions of real objects tend to differ from the actual physical form of the object in many critical ways as exemplified by humans' tolerance of spatial inconsistencies and the prevalence of robust spatial and kinetic illusions. Such phenomena have been suggested to be byproducts of complicated mathematical computations or errors in neural mapping. Here, we show that a basic parallel distributed processing network that was trained to recognize shapes is susceptible to illusory contour-style illusions. The network learned to identify shapes accurately and robustly given only unambiguous sensory input. When tested with Kanizsa triangle-type illusions, the network made systematic "errors" when identifying shapes that were consistent with human visual illusions. Shape recognition occurs when various features of an image have been processed by distributed networks that seem to specialize in certain types of features (e.g., edges, angles, colors, moving edges). Higher centers then "recognize" when stimuli from feature processing areas reach a critical mass. Our data suggest that the distributed nature of the information-processing in our network may also contribute to "machine illusions." Natural stimuli tend to be ambiguous, and this ambiguity is resolved by matching the actual stimuli with known images based on past visual experience. Since resolving stimulus ambiguity is a challenge faced by all visual systems, a corollary of these findings is that humans might experience illusions because of a similar mechanism. The data might also provide an alternate definition of illusion: The extent to which a true stimulus matches the stored image which the stimulus most likely appears to be. Accordingly, illusions may not be fundamentally different from non-illusory percepts, both being direct manifestations of the relation between images and what the images are expected to be.
Meeting abstract presented at VSS 2013