Home Research Research Highlights
O'Reilly Workshop: Models of High-Level Vision
Members of the Interacting Memory Systems Network and the Perceptual Expertise Network held a two-day workshop on models of high-level vision at the University of Colorado at Boulder on July 21 and 22, 2008. Participants included Randy O’Reilly of IMSN and Tim Curran, Tom Palmeri, and Gary Cottrell of PEN and several members of their laboratories. This intensive workshop focused on general issues surrounding high-level vision and object recognition. The workshop featured a discussion of the highvis model that has been in development by the O’Reilly laboratory for several years with a publication in preparation. The overall approach of the highvis model to solving the object recognition problem is similar to many existing models by Fukishima, Riesenhuber and Poggio, Mozer, Deco and Rolls, and others, with a hierarchy of increasing spatial invariance and featural complexity. A major feature that distinguishes this model from most previous models is that it includes both feedforward and feedback connections, there is a learning mechanism (based on the Leabra learning rules), and there is a spiking implementation. The implementation also does well at standard object and face recognition tasks. The workshop featured talks by the attendees but concentrated most of the time on a lot of in depth discussion about highvis in particular and models of high-level vision in general.