From University of Pennsylvania
Scientists receive $3 million to develop biologically-based artificial vision systems
PHILADELPHIA -- A team of researchers from three universities, led by a University of Pennsylvania bioengineer, has won a $3 million grant for work toward artificial-vision technologies that might detect patterns as robustly as the human brain. The work could lead to satellite-based means of detecting environmental destruction, automated systems to detect abnormalities in mammograms and other medical images and computerized approaches to other tasks now possible only through the discretion of the human eye.
The five-year award comes via the Multi-University Research Initiative at the Office of Naval Research, which hopes to gain a means of better integrating infrared, visual and ultraviolet images from satellites. Of 48 MURI awards this year, the Penn-led effort is one of just two in neuroengineering, uniting bioengineers and neuroscientists.
The project aims to move beyond the current limitations of computer simulations of the brain's visual cortex, which have proven inept at the kinds of generalization and pattern recognition that underlie intelligence. Unlike a person, a state-of-the-art artificial neural network that has "seen" hundreds of different chairs often encounters difficulties identifying a new chair as part of the same category.
Principal investigator Leif H. Finkel, Penn professor of bioengineering, said that enabling automated systems to recognize such visual patterns could eventually take the pressure off skilled clinicians to scan endless medical images for irregularities. Mounted in satellites, the technology could survey patterns of land use or monitor the transformations wrought by global climate change.
Finkel will lead a team of researchers from Penn, Columbia University and the Massachusetts Institute of Technology. The group hopes to surmount artificial neural networks' shortcomings, making them more like the human brain's visual cortex in their ability to recognize relationships among factors. Such "smart" networks could accurately infer patterns and correlations even when supporting information is limited.
Finkel's collaborators on the project include Penn's Kwabena Boahen, assistant professor of bioengineering, and Diego Contreras, assistant professor of neuroscience; Paul Sajda, associate professor of biomedical engineering at Columbia; and Edward Adelson, professor of brain and cognitive science at MIT.
Contreras will provide key data for building these models through his recordings from the visual cortex. Boahen will develop novel hardware for the project, such as VLSI chip-based neural networks. Sajda will develop the fundamental mathematical underpinnings and bridge the probabilistic analysis to medical and other image applications. Adelson, a renowned expert in human psychophysics and visual processing of motion, will help detect patterns in real-time systems subject to movement. The project also involves interactions with several government laboratories and private corporations.
At Penn, the award will support four graduate students and one postdoctoral researcher annually.