Prof Thomas Wallis (TU Darmstadt) - Artificial neural networks as approximate upper bounds on performance

Abstract

When trying to understand a system like the human mind, cognitive scientists usually ask people to perform a task. Measuring how people behave given a certain task allows us to understand the processing strategies used and limitations that are present. How well do hypothesised mechanisms explain behaviour? To answer this question, it helps if we can understand the upper bounds: how well could you explain the data? How well could a system perform this task? Ideal observers can directly answer these questions, but only for limited tasks. I will argue that artificial neural networks can be a useful tool to place approximate upper bounds on performance in settings where traditional methods cannot be applied. As examples, I will present work on predicting eye movement behaviour during scene viewing, and discriminating closed from open contours in cluttered displays.

Bio

I am the Professor for Perception at the Institute of Psychology and Centre for Cognitive Science, TU Darmstadt. My interests mostly focus on visual perception in humans and machines, but also include machine learning and cognitive modelling and applications of visual perception research (image quality, driving).

I finished my PhD in Psychology in 2010 in Derek Arnold's lab at the University of Queensland, Australia, followed by a postdoctoral fellowship with Peter Bex in Boston, before moving to Germany in 2013 to work with Felix Wichmann and Matthias Bethge. I spent two years working as a scientist at Amazon research before joining the TU Darmstadt as a professor in 2021.