April 28, 2014, 04:15 pm – 05:45 pm

Location: S2 02|C110 Hochschulstrasse 10 64289 Darmstadt

Speaker: Prof. Constantin A. Rothkopf, PhD

Here we will  present several studies that utilize a combination of machine learning, classic behavioral experiments, and virtual reality technology to demonstrate the necessity of considering perception, action, and learning jointly in the quantitatively description of human behavior in extended sequential naturalistic tasks.
First, work on quantifying human sequential behavior in terms of underlying costs and benefits in the framework of inverse optimal control will be presented leading to a description and prediction of natural navigation behavior.
Secondly, using a virtual environment allowing for the manipulation of the statistical relationship between observation uncertainties and behavior, it will be shown that the interception of moving targets that classically has been described as a fixed rule heuristic is instead highly adaptive so as to allow human subjects to increase their interception probability.
Finally, it will be shown, that representational learning ideas cannot explain several properties of cortical neurons on the basis of natural image statistics alone but require taking into account the statistics of the natural environment, the statistics imposed by the visual system, and the statistics of active visual behavior.