Jeremy Howard. Jon Snyder/WIRED
Some of the world’s most brilliant minds are working as data scientists at places like Google, Facebook, and Twitter—analyzing the enormous troves of online information generated by these tech giants—and for hacker and entrepreneur Jeremy Howard, that’s a bit depressing.
Howard, a data scientist himself, spent a few years as the president of the Kaggle, a kind of online data scientist community that sought to feed the growing thirst for information analysis. He came to realize that while many of Kaggle’s online data analysis competitions helped scientists make new breakthroughs, the potential of these new techniques wasn’t being fully realized. “Data science is a very sexy job at the moment,” he says. “But when I look at what a lot of data scientists are actually doing, the vast majority of work out there is on product recommendations and advertising technology and so forth.”
So, after leaving Kaggle last year, Howard decided he would find a better use for data science. Eventually, he settled on medicine. And he even did a kind of end run around the data scientists, leveraging not so much the power of the human brain but the rapidly evolving talents of artificial brains. His new company is called Enlitic, and it wants to use state-of-the-art machine learning algorithms—what’s known as “deep learning”—to diagnosis illness and disease.
His basic idea is to create a system akin to the Star Trek Tricorder, though perhaps not as portable.
Publicly revealed for the first time today, the project is only just getting off the ground—”the big opportunities are going to take years to develop,” Howard says—but it’s yet another step forward for deep learning, a form of artificial intelligence that more closely mimics the way our brains work. Facebook is exploring deep learning as a way of recognizing faces in photos. Google uses it for image tagging and voice recognition. Microsoft does real-time translation in Skype. And the list goes on.
But Howard hopes to use deep learning for something more meaningful. His basic idea is to create a system akin to the Star Trek Tricorder, though perhaps not as portable. Enlitic will gather data about a particular patient—from medical images to lab test results to doctors’ notes—and its deep learning algorithms will analyze this data in an effort to reach a diagnosis and suggest treatments. The point, Howard says, isn’t to replace doctors, but to give them the tools they need to work more effectively. With this in mind, the company will share its algorithms with clinics, hospitals, and other medical outfits, hoping they can help refine its techniques.
Deep-Learning Doctors
Howard says that the health care industry has been slow to pick-up on the deep-learning trend because it was rather expensive to build the computing clusters needed to run deep learning algorithms. But that’s changing.
Jeremy Howard and senior data scientist Choon Hui Teo look at some of the latest research in deep learning for detection of mitotic activity for detecting breast cancer. Enlitic
Howard isn’t the only one exploring these possibilities. He says academic researchers such as Stanford computer scientist Daphne Koller have already made progress in applying deep learning to medicine. And then there’s IBM, whose Jeopardy-winning supercomputing system, Watson, is using machine learning to aid doctors at New York’s Memorial Sloan-Kettering Cancer Center.
But Watson doesn’t use deep learning per se—it uses older techniques—and Howard says the overall approaches taken by two companies are very different. IBM is essentially feeding Watson medical text books in an attempt to teach it what doctors already know, he says, while Enlitic is feeding the raw data into its machines, letting the computers find the patterns between certain symptoms and treatments with different outcomes. In other words, Watson mimics medical science in the pursuit of creating a artificial super doctor that knows more than any single doctor could ever learn. But Enlitic could potentially make new discoveries by uncovering previously unnoticed patterns in the data.
The Real Challenge
The real challenge, Howard says, isn’t writing algorithms but getting enough data to train those algorithms. He says Enlitic is working with a number of organizations that specialize in gathering anonymized medical data for this type of research, but he declines to reveal the names of the organizations he’s working with. And while he’s tight-lipped about the company’s technique now, he says that much of the work the company does will eventually be published in research papers.
Even with expert help, trying to create such a system is an intimidating task. After all, the hope is that people will trust their lives to Enlitic. “Certainly, we’re doing something more risky than giving someone a product recommendation they didn’t like,” Howard says. But he’s undaunted. After all, the potential reward is far greater.
No comments:
Post a Comment