Departmental Colloquium: Marc Ernst - University of Bielefeld - From Multisensory Perception to Sensorimotor Behaviour: A Probabilistic Approach

Abstract: The human brain uses multiple sources of sensory information together with prior knowledge about the statistical regularities of the world in order to generate the most appropriate action. However, there is uncertainty due to noise and ambiguity in the sensory information. Furthermore, sensory information as well as prior knowledge may be imprecise and possibly inaccurate for the current action context. Therefore, a decision for choosing some action can only be taken probabilistically. Bayesian Decision Theory, which I will review here in this talk, provides a suitable probabilistic framework to come up with ideal observer models, against which human perception and performance can be compared. Besides an introduction to this general framework, I will talk about two of our recent studies in more detail, which are examples of the application of this framework to multisensory perception and action. The first study deals with the decision process, which signals to combine. Sometimes there are seemingly arbitrary association between multisensory signals such as, for example, sound frequency and spatial elevation. This goes so far that even in most natural languages the spatial labels �high� and �low� are used for particular sound frequencies. We recently showed (Parise et al., PNAS, 2014) the reason for this association to be found in the statistics of the natural environment where high frequency sounds are more likely to originate from a higher spatial elevation and vices versa. The second study (van Dam, Ernst, PLoS One, 2013) investigates the knowledge we have about the errors we make when executing actions. This is important because it is the knowledge about such errors that shapes sensorimotor learning. We show that we have a surprisingly detailed understanding of even the random errors we make and that we can use this knowledge in a statistically optimal way for the correction of the errors. I will end my talk with a discussion about the benefits and the limits of this framework.