Colloquium Cognitive Systems
Object recognitionin man and machine
Prof. Dr. Felix Wichmann (University of Tubingen)
Abstract. Convolutional neural networks (CNNs) have been proposed as computational models for (rapid) human object recognition and the (feedforward-component) of the primate ventral stream. The usefulness of CNNs as such models obviously depends on the degree of similarity they share with human visual processing. In my talk I will discuss two major differences between human vision and currently used standard CNNs: First distortion robustness---CNNs fail to cope with novel, previously unseen distortions. Second texture bias---unlike humans, standard CNNs primarily recogniseobjects by texture rather than shape. However, both differences between humans and CNNs can be overcome: we created a suitable dataset which induces a human-like shape bias in CNNs during training. This resulted in an emerging human-level distortion robustness in CNNs. Taken together, our experiments highlight how key differences between human and machine vision can be harnessed to improve CNN robustness by inducing a human-like bias---and thus make CNNs more similar to the human visual system.
Bio. Felix Wichmann received his B.A. (1994) and D.Phil. (1999) in Experimental Psychology from the University of Oxford. After post-doctoral research at the University of Leuven (2000-2001), he worked as a research scientist in the EmpiricalInference Department at the Max Planck Institute for Biological Cybernetics in Tübingen (2001-2007). From 2007 to 2011 he was Associate Professor (W2) at the Technical University of Berlin and since 2011 he is Full Professor (W3) at the Eberhard Karls Universität Tübingen. He serves on the editorial board of the Journal of Vision.