Abstract
Illusions provide a great opportunity to study how perception is affected by both the observer’s expectations and the way sensory information is represented1,2,3,4,5. Recently, Jazayeri and Movshon6 reported a new and interesting perceptual illusion, demonstrating that the perceived motion direction of a dynamic random dot stimulus is systematically biased when preceded by a motion discrimination judgment. The authors hypothesized that these biases emerge because the brain predominantly relies on those neurons that are most informative for solving the discrimination task7, but then is using the same neural weighting profile for generating the percept. In other words, they argue that these biases are “mistakes” of the brain, resulting from using inappropriate neural read-out weights. While we were able to replicate the illusion for a different visual stimulus (orientation), our new psychophysical data suggest that the above interpretation is likely incorrect: Biases are not caused by a read-out profile optimized for solving the discrimination task but rather by the specific choices subjects make in the discrimination task on any given trial. We formulate this idea as a conditioned Bayesian observer model and show that it can explain the new as well as the original psychophysical data. In this framework, the biases are not caused by mistake but rather by the brain’s attempt to remain ’self-consistent’ in its inference process. Our model establishes a direct connection between the current perceptual illusion and the well-known phenomena of cognitive consistency and dissonance8,9.
