Artificial Vision on Par with Human Eyesight

A team of researchers from three universities is working on artificial vision technologies that could one day detect visual patterns as effectively as the human brain. Funded by a $3 million grant from the U.S. Office of Naval Research, the work may eventually lead to satellite-based means of detecting environmental destruction, automated systems that find abnormalities in medical images and computerized approaches to other visual tasks now possible only through the discretion of the human eye.

By combining the work of bioengineers and neuroscientists, the project aims to improve on today’s rudimentary computer simulations of the brain’s visual cortex. That would be a big step forward, since current state-of-the-art artificial neural networks are unable to handle the kinds of pattern recognitions that underlie visual intelligence. Unlike a furniture expert, for example, who can deftly categorize hundreds of different chairs, even the most sophisticated neural network would have trouble sorting the chairs it “saw” into specific categories.

Designing automated systems to recognize visual patterns could eventually take the pressure off skilled clinicians, who currently must scan endless medical images in order to seek out irregularities, says Principal Investigator Leif H. Finkel, a University of Pennsylvania professor of bioengineering. Mounted in satellites, the technology could survey patterns of land use or monitor the transformations wrought by global climate change.

Finkel’s collaborators include University of Pennsylvania colleagues Kwabena Boahen, an assistant professor of bioengineering, and Diego Contreras, an assistant professor of neuroscience. Paul Sajda, an associate professor of biomedical engineering at Columbia University, and Edward Adelson, professor of brain and cognitive science at MIT, are also on the team. Contreras is providing key data for neural model construction based on his recordings of the visual cortex. Boahen is developing the project’s hardware, including chip-based neural networks. Sajda is creating the project’s mathematical underpinnings. And Adelson, an expert in human psychophysics and visual motion processing, is developing models that detect moving patterns in real-time systems. The researchers hope to report their progress, if any, in a couple of years.