Despite the dizzying array of smart technologies now defining modern life, at a deep level the programs running these gadgets come in two basic kinds: those that can steer a car and those that can finish a sentence. But in seeking smarter medical technologies (those that can detect a tumor), Princeton researchers have developed a third way forward that makes artificial intelligence work better for human health.
Ozge Akmandor, graduate student in electrical engineering and first author on a paper outlining the new approach, said her method combines the best of both conventions to address major healthcare problems like disease monitoring and prevention. In October, Akmandor won first place in three out of three categories at the Internet of Medical Things (IoMT) competition, held in conjunction with a broadbased series of events in New York City called Embedded Systems Week.
The innovation changes how machines classify information, a key first step in predicting and responding to worldly conditions. Conventional machine-learning algorithms classify sensor data based on either features (as with image processing) or semantics (as with natural language processing). Features can be anything from the color of a pixel to the pattern of an electronic pulse. Semantics deals with the meanings of and relationships between words. The new approach does both.
"Since the feature and semantic spaces look at the classification problem from complementary angles, their combination leads to a significant boost in accuracy," said Niraj Jha, professor of electrical engineering, who is Akmandor's adviser and the study's principal investigator.
Devices like home security systems, activity-monitoring smart watches, autonomous vehicles and music playlist generators all take advantage of what experts call the feature space. These programs, for example, resolve pixels into stop signs and electrical pulses into heart rates. Other devices work in the semantic space, converting speech into type or translating text between languages. In complex systems, these two types of algorithms might contribute separately to a seamless overall experience, as with virtual assistants created by Apple, Google and Amazon. But no one had yet leveraged the power of both spaces in a single algorithm.
Akmandor and her team found a way. In the IoMT competition in October, their approach led in all three performance categories—accuracy, time and energy use—the three key metrics engineers use to evaluate an algorithm. The results delighted Akmandor, who had focused primarily on accuracy in her design. Greater precision often comes with trade-offs: slower or more energy-intensive computations. But the new method, called Semantically Enhanced Classification of Real-World Tasks, or SECRET, achieved a rare win-win-win scenario, improving all three metrics together.
The idea for SECRET originated when Akmandor began as a research intern at IBM's Thomas J. Watson Research Center. She continued refining the tool at Princeton under Jha's advising. Additional contributors include Jorge Ortiz, formerly at IBM and now an assistant professor at Rutgers University; and IBM researchers Irene Manotas and Bong Jun Ko.
Their work points the way toward new directions in how machines, employed smartly, can revolutionize the way we manage human health.