Published On January 27, 2015
Artificial intelligence in scientific research—like AI in medicine—has a long but so far underwhelming history. In the early 1970s, enthusiasts predicted that computers would soon play a far more active role in science, progressing from crunching numbers to reasoning side-by-side with researchers. Although that hasn’t happened yet, a new generation of intelligences is pushing into new frontiers.
Eureqa, the best-known of these, was developed by Cornell University roboticist Hod Lipson and graduate student Michael Schmidt, and debuted in 2009 in a study published in the journal Science. Starting from scratch with raw data generated by a swinging pendulum, the program inferred Newton’s second law of motion—that force equals mass multiplied by acceleration— and the law of conservation of energy.
To analyze the data, Eureqa used algorithms inspired by programs Lipson wrote to make self-repairing robots. Those robots had to identify problems, design a fix and evaluate the results; they were, in a sense, autonomous. Given scientific form in Eureqa, the approach was well-suited to handle a post-millennial explosion of extraordinarily complicated genetic and biochemical data. Just plug in the information, all those formless gigabytes of protein activity and interacting genes, and Eureqa discovers equations describing the underlying dynamics.
Sometimes, Eureqa—which now boasts more than 40,000 users, many working in the life sciences–even finds equations that its users don’t know how to explain. These are, in a sense, answers in search of a hypothesis, and underscore a basic AI dynamic: computer insights lead inevitably to more questions, and human thought still plays a central role in making sense of it all.
Computers might become more helpful in that process, though. In a recent Science article, Jim Hendler, an AI researcher and director of the Institute for Data Exploration and Applications at Rensselaer Polytechnic Institute, described a new wave of AIs that combine data-processing abilities of the sort contained in Eureqa with language skills, image analysis and models of scientific reasoning. The result: an “artificially intelligent assistant” that thoughtfully digests scientific studies, data from experiments and even online discussions.
“It’s not going to say, ‘Here is a paper with these keywords in it,'” says Hendler. “It’s going to say, ‘Here are the different ideas coming out of these documents. If you read this one, you’ll understand this whole other set of results; and this one might be most relevant to what you were working on yesterday.”
These high-level AIs could even suggest experiments and review data in light of a scientist’s own work, says Hendler, who is collaborating with RPI biomedical engineer Juergen Hahn and other faculty to build such an assistant, based on a biological system Hahn is modeling. IBM has also fashioned a version of Watson, their Jeopardy!-winning AI, called Watson Discovery Advisor; it analyzes scientific evidence and suggests avenues for future research, such as sets of proteins that interact with a gene linked to cancer. Watson is also being used to speed up the analysis of cancer clinical trial data and match cancer patients with clinical trials.
“There are some things humans are good at, and some things computers are good at,” says Hendler. “And neither one alone is as good as the two together.”
Stay on the frontiers of medicine