Researchers from Harvard and MIT have demonstrated a way to build better artificial visual systems with the help of low-cost, high-performance gaming hardware.
Using graphics processing units, or GPUs — the same technology that makes your favorite video game possible to play with detailed visual graphic effects — along with a screening technique borrowed from genetics, the team is attempting to mimic the human brain’s ability to process visual data.
The amount of computational crunching your brain does to recognize a basic object is “profoundly difficult to mimic,” according to the researchers. A biological visual system has hundreds of millions of processing units, according to the researchers, and the challenge is more than just massing all that horsepower in one place.
The study, co-led by David Cox of the Rowland Institute at Harvard and Nicolas Pinto of the McGovern Institute for Brain Research and the Department of Brain and Cognitive Sciences at MIT, is an attempt to figure out how all that biological equipment works together by building a computing system with actual computer equipment.
“While studying the brain has yielded critical information about how the brain is wired, we currently don’t have enough information to build a computer system that works like the brain does,” Pinto said in prepared remarks. “Even if we take all of the clues that we have available from experimental neuroscience, there is still an enormous range of possible models for us to explore.”
That’s where the genetic screening technique comes into play. Molecular biologists screen many candidate organisms or compounds in parallel to find those with a particular property of interest.
Expanding on that technique, the team constructed thousands of candidate models and screened for those that performed best on an object recognition task.
The resulting models “outperformed a crop of state-of-the-art computer vision systems across a range of test sets,” according to the researchers, and more accurately identifyied a range of objects on random natural backgrounds with variation in position, scale and rotation.
With run-of-the-mill computer processing units (CPUs), the effort would have required years of time or computing hardware worth millions of dollars.
With modern, GPU-oriented gaming hardware, the analysis was done in a single week for considerably less expense.
“GPUs are a real game-changer for scientific computing,” Pinto said in his remarks. “We made a powerful parallel computing system from cheap, readily available off-the-shelf components, delivering over hundred-fold speed-ups relative to conventional methods. With this expanded computational power, we can discover new vision models that traditional methods miss.”
The team’s approach could yield many potential applications, including face identification, object tracking, gesture recognition and pedestrian detection for automotive applications.
Their results were published in the Nov. 26 issue of the journal PLoS Computational Biology.
