Investigating Human Perception with Neurocomputational Models

Sponsored by the NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai

What is the underlying basis of visual perception? There are many approaches to this question: behavioral studies, EEG, fMRI, MEG, recordings from animals, and recordings from humans being operated on for epilepsy. Another approach is to build computational models that “do the same things humans do,” in a brain-like way, i.e., using neural networks. Given a working model of some task, it can be manipulated and analyzed in ways that are impossible in humans or animals. A working model often can account for a wide range of data, as well as make predictions for future research.

In this talk I will review a number of models that can provide insights into visual perception. I will begin with an efficient coding model (our version of sparse PCA), that learns center-surround receptive fields when exposed to natural images, magno- and parvo-like representations when exposed to video, and gammatone filters when exposed to sound. Thus the model can explain peripheral representations in more than one modality. I then move to a higher level model that can explain multiple results in face and object perception, in particular, why the Fusiform Face Area becomes recruited for other objects of expertise. My conclusion here is that there is nothing special about faces per se, it is what we have to do with them - that is, fine-level discrimination - that makes them special.

Biography

Garrison W. Cottrell is a Professor of Computer Science and Engineering at UC San Diego. He is Director of the Interdisciplinary Ph.D. Program in Cognitive Science at UCSD, and the Director of the Temporal Dynamics of Learning Center, an NSF-sponsored Science of Learning Center involving 40 PIs at 18 institutions in four  countries. He is also a founding PI of the Perceptual Expertise Network. 

Professor Cottrell's main interest is Computational Cognitive Neuroscience, in particular, building working models of cognitive processes and using them to explain psychological or neurological processes. In recent years, he has focused upon unsupervised feature learning (modeling precortical and cortical coding), face & object processing, visual salience, and visual attention. He has also worked in the areas of modeling  psycholinguistic processes, such as language acquisition, reading, and word sense disambiguation. He received his PhD. in 1985 from the University of Rochester under James F. Allen (thesis title: A connectionist approach to word sense disambiguation). He then did a postdoc with David E. Rumelhart at the Institute of Cognitive Science, UCSD, before joining the CSE Department in 1987.