Colloquium Cognitive Systems

Perceiving motion at twilight: Lights out for the Bayesian observer?

Prof. Tom Freeman ( Cardiff University )


Abstract. There are myriad ways we can change perceived speed. Driving in fog makes motion look slower, small things move faster than big things, blue travels more slowly than red, eye tracking slows moving targets down. ‘Mechanistic’ theories emphasise the inaccuracy of early motion sensors – initial motion measurements depend on stimulus context. But why should the brain accept such errors, why aren’t they calibrated out? ‘Bayesian’ theories provide a more principled approach. They consider perceptual estimates as the result of a statistical ‘best-guess’ in the face of sensors with variable output. Changes in perceived speed are therefore a statistical compromise between sensor imprecision, which can change with visual context, and a prior expectation that the world is largely stationary. In my talk, I will first show the potential of the Bayesian framework to provide a general account of motion perception, one that cuts across different types of motion sensor (not just those to do with the processing of visual images) and different types of individuals. I will then explore a visual context that appears to challenge a crucial assumption of the Bayesian approach, namely that early sensors are accurate. As light levels fall, motion appears to speed up, which seems counter to the Bayesian approach: surely sensor output should be more variable at low light levels, and so motion slow down? I will show that this is not the case: sensors are never more variable in the dark, they are either less variable (surprise!), or about the same as in the light, once perceived contrast is controlled for. In all cases, however, motion appears just as fast in the dark (surprise!). Further experiments provide a strong case for a hybrid Bayesian model, one in which prior expectations exert an influence on perceptual estimates, even when the accuracy of initial sensing depends on contexts like prevailing light level.

Bio. Tom Freeman is a Professor at the School of Psychology at Cardiff University, where he is the Director of Postgraduate Research. His main research interests are in perception and self-motion, such as the role of eye movements in the perception of visual motion and depth, and more recently how head movements influence the perception of auditory space. The underlying approach uses psychophysics and the Bayesian framework. He carried out his PhD at the University of Birmingham on the visual perception of optic flow with Mike Harris, and then post-docs with Glyn Humphreys on connectionist networks, Mark Georgeson at Aston University on the perception of edge location and blur, and Marty Banks at the University of California, Berkeley, on perception of self-motion and object motion during smooth pursuit eye movement. Recent collaborations include work on vestibular motion perception with Marc Ernst (Ulm University), and auditory motion processing with David Alais (University of Sydney), schizophrenia and spatial vision with Krish Singh (Cardiff University), Rosalyn Moran (University of Bristol) and Karl Friston (UCL).