As artificial intelligence becomes smarter, it has the potential the radically transform the practice of scientific discovery, promising a new era of breakthroughs on complex problems like global warming and personalized genomic medicine. How? By making science less error prone, easier to reproduce, and by getting around pesky human bottlenecks that slow down discovery.
That’s the argument made in a article today by four researchers–University of Southern California’s Yolanda Gil, Pacific Northwest National Laboratory’s Mark Greaves, RPI’s James Hendler, and Cornell University’s Haym Hirsh–published in the journal Science.
Using AI in science is not new, the authors note, and dates at least back to the 1970s when Nobel laureate Herbert Simon developed systems with the ability to reason about scientific data. But today such systems are advancing at rapid speeds in helping researchers make predictions about large data sets that are impractical to process at human speed. One example given by the researchers is IBM’s Watson. Among other uses in medicine, it is being deployed to not only help doctors and researchers process tumor genomes in brain cancer patients, but to help them understand what personalized treatments to offer.
Newer systems in development are also going a step beyond. “These systems enhance the intelligent assistants described earlier with the capability to attack scientific tasks that combine rote work with increasing amounts of adaptivity and freedom,” the paper says, allowing them to “assist in tasks that previously required human knowledge and reasoning.”
Examples include the “Hanalyzer,” finds academic papers that are relevant to a specific scientist’s work and is able to make new correlations, suggesting, for example, new genes that are promising to investigate. Another system, Eureqa, is practically a scientist all on its own: it “searches a vast space of hypotheses consistent with given data observed in an experiment, selects those most promising, and designs experiments to test them.” A system called Sunfall is able to reduce scientists’ workload and rate of false positives as they are identifying supernova.
“When successful, the computer can become a real (although junior) participant in the science process doing what it does best: applying algorithmic models and bringing knowledge to bear in a consistent, systematic, and complete manner.”
Despite this potential, many scientists appear reluctant to embrace the use of AI–in part because of lack of training and outreach from the AI field. Though AI software, from speech recognition to autonomous cars, is very present in the consumer world, it is less enthusiastically embraced by the research community. The authors call on this to change.
They write: “The world faces deep problems that challenge traditional methodologies and ideologies. These challenges will require the best brains on our planet. In the modern world, the best brains are a combination of humans and intelligent computers, able to surpass the capabilities of either alone.”