In 2011, Artificial Intelligence (AI) came of age when IBM’s Watson computer beat two human contestants to win Jeopardy. These were not any two contestants. Ken Jennings had won 74 times consecutively and Brad Rutter had pocketed the biggest pot in history - $3.25 million. In the battle between man and machine, Watson’s win was historic. Jennings was sanguine about losing: “I, for one, welcome our new computer overlords.”
The key to Watson’s success is a technique called “deep learning”, which has achieved astonishing results in several domains, most notably in understanding natural language. Research teams use it to teach computers to find meaning in vast amounts of text. Disrupting quiz shows is one thing, but can AI disrupt science?
Perhaps we should first ask, does science need disrupting? Yes. Access to reliable knowledge – the academic literature – is becoming a fundamental bottleneck for humanity. There are now over 50 million research papers and this is growing at a rate of over one million a year. Over 70,000 papers have been published on a single protein – the tumor suppressor p53. How can any academic keep up? And how can anyone outside of academia make sense of it all - the public, policy makers, business people, doctors or teachers? Well, most academics struggle and the public can’t – most research is locked behind pay walls.
Ironically, in the age of the internet and unparalleled access to information, the most critical is out of bounds. Moreover, while we are clearly pretty good at producing knowledge, using this knowledge – that is separating the wheat from the chaff and integrating this together into something useful – is a big problem particularly in fields such as global sustainability.
That may be about to change. Here are five ways AI looks set to disrupt science.
1. Science mining #1: Iris.AI
The co-founders of this new AI start-up believe that if you could access and contextualize all of the world’s published research you would solve a lot more problems. So they’ve set out to do that. “We want to democratize access to scientific knowledge. The first step is a science assistant leveraging AI and the crowd to help users map out and find relevant scientific knowledge,” says Finnish co-founder Maria Ritola. The team has created an AI tool for innovators to do quick mappings of a research area, but in the long-term they want to build an AI scientist that can create a hypothesis based on existing publications, run experiments and simulations and even publish papers on the results. So they are not short of ambition.
They started with a simple tool to map out the science around a TED talk. Iris analyses the scripts of the talks using Natural Language Processing algorithms, mines open-access academic literature to find key papers related to the talk’s content, then visualizes, really quite beautifully, the groups of related research papers. “Iris is a young AI. We call her our baby. She doesn’t get everything right yet - she’s at about 70% accuracy - so we’re enlisting a community of AI Trainers to help her learn,” Ritola explains. The TED tool is only the tiny first step, and in September the Iris team will launch the first commercial tool with a group of corporate R&D departments as pilot users.
2. Science mining #2: Semantic Scholar
This is genius. Or will be one day. Semantic Scholar is an academic search engine from Microsoft co-founder Paul Allen’s Allen Institute for Artificial Intelligence. It too uses AI to search the academic literature and it is impressively fast. Still in beta, it’s interface is less elegant, more “academic utilitarian” than the sleek Iris. And it also throws up some oddities. A search for the landmark paper “A safe operating space for humanity”, which appeared in Nature (along with Science, one of the two leading academic journals) in 2009, does not show up on the first page, nor does the follow up paper, which appeared in Science in 2015. The full-length version of the original paper, published in the more obscure journal Ecology and Society, is, however, the first entry. But, given the team only started in 2015, and they are first focusing on computer science papers, this is not bad going.
The team is adding new features. It can already trawl through a paper’s reference list and work out which citation has been genuinely influential and which is just background.
3. From miner to scientist
A team at IBM and colleagues has gone a step further than both Iris.AI and Semantic Scholar. They say their system can do science. That is, their AI can generate scientific hypotheses automatically by mining academic literature. Moreover, their algorithms, they say, can be used to make new scientific discoveries. Their goal is to combine text mining with visualization and analytics to identify facts and suggest hypotheses that are “new, interesting, testable and likely to be true”, the authors say in a 2014 research paper.
4. Science media: Science Surveyor
Science journalists are the target for this bit of AI from a collaboration between Columbia and Stanford universities. Science Surveyor has been designed to help journalists assess the significance of a new piece of research. Where does the research fit into the bigger picture in the field? Is it really ground-breaking? Is it contested? Journalists need answers to all these questions on impossible deadlines and often with little expert knowledge in the area. This leads to churnalism, poor reporting and a readership that is none the wiser. Science Surveyor is an interesting experiment to move beyond this.
5. Open Access AI: Open.ai
Sponsored by PayPal founders Elon Musk and Peter Thiel, among others, OpenAI is a non-profit research company that aims to democratize AI. "It's really just trying to increase the probability that the future will be good," Musk said at the recent Recode Code Conference. Its goal “is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.” This is a wildcard. No one is really sure what could come out of it, but if Musk is behind it, expect the unexpected – fast.
AI could be on the cusp of driving the next phase of the scientific revolution, yet this is less discussed than the bigger existential threat of AI. In many ways, AI innovations could simply help scientists to do their jobs more efficiently – thereby cutting the crippling time lag between science and society. For example, could machine learning algorithms delve deep into the previous five assessment reports of the Intergovernmental Panel on Climate Change and, based on research published since the last report, provide rudimentary conclusions of the sixth report?
One major hurdle to progress is the academic publishers. They have no financial incentives to grant AI initiatives access to the body of human knowledge, though Google Scholar has permission to trawl all text behind paywalls, so this may not be insurmountable.
We have launched the Future Earth Media Lab to break down the barriers between science and society. Lab co-founder, and author of this piece, Owen Gaffney is a mentor for the Iris.AI project. Our mission is to seed, nurture and develop similar initiatives that can – project by project – nudge the technological revolution towards better outcomes for science and society, ultimately towards sustainable futures.