The application of AI to science (sometimes termed ”scientific discovery”) is of both philosophical and economic importance. Its philosophical importance is that if the scientific process could be automated it would demonstrate, in an operational sense, that we understand the scientific process; at the time of his death, the famous physicist Richard Feynman had written on his blackboard, ”What I cannot create, I do not understand”. Its economic importance is that the application of AI to science has the potential to increase the productivity of scientific research. And, as science is the greatest driver of economic growth, more productive scientific research will result in faster economic growth.

AI systems can increase the productivity of science by augmenting human scientific reasoning in multiple ways. They are capable of superhuman reasoning abilities and are able to:

• flawlessly remember vast numbers of facts
• execute flawless logical reasoning
• execute near optimal probabilistic reasoning
• learn more rationally than humans from small amounts of data
• learn from large amounts of data which no human could deal with
• extract information from millions of scientific papers

The integration of an AI system with laboratory automation is termed a “robot scientist”. A robot scientist can automatically originate hypotheses to explain observations; devise experiments to test these hypotheses; physically runs the experiments using laboratory robotics; interpret the results to change the probability of hypotheses; and then repeat the cycle. Robot scientists allow:

• Faster scientific discovery. Robot scientists can generate and test thousands of hypotheses in parallel (utilising experiments that test multiple hypotheses). Human cognitive limitations mean that they can only consider a few hypotheses at a time
• Cheaper experimentation. Robot scientists can be designed to select experiments with more economic rationality – saving time and money
• More easily reproduced and trained. To train a human scientist requires over 20 years and a huge amount of resources. Humans can only absorb knowledge slowly through teaching and experience; robot scientists can directly absorb knowledge from each other
• Harder working. Robots can work longer and harder than humans. They work 24/7 and do not require rest or holidays
• Improved knowledge/data sharing, and reproducibility. One of the most important current issues in biology is ”the reproducibility crisis”: “There is growing alarm about results that cannot be reproduced. Explanations include increased levels of scrutiny, complexity of experiments and statistics, and pressures on researchers”. Robot scientists have the superhuman ability to record experimental actions and results. Their results, along with the associated meta-data and employed procedures, are automatically recorded and deposited in full, and in accordance with accepted standards, at no additional cost. Moreover, despite the recording of experimental data being widespread, it is still uncommon to fully document used procedures, errors and all the meta-data.

One of the paradoxes of AI is that it is not intuitively obvious what human tasks are hard to automate. For example, it is very surprising that it was far easier to develop a robot to beat the world chess champion, than a robot capable of walking across a road.

Science is arguably the most profound creation of the human mind, yet it has attributes that favour the application of AI compared to messier domains. Like chess, scientific problems are abstract but involve the real-world and they are are restricted in scope – no need to know about “Cabbages and Kings” (Lewis Carroll in Through the Looking Glass). Nature is honest – we may misunderstand some experimental results, but the real-world is not conspiring to confuse us. This makes science much easier for AI systems than tasks which involve dishonest human agents.

In chess there is a range of ability from novices up to Grandmasters. I argue that this is also true in science, from the simple research of robot scientists now to what most human scientists can achieve, all the way up to the ability of a Newton or an Einstein. If you accept this then, just as in chess, it is likely that advances in computer hardware and software will drive the development of ever-smarter robot scientists. In favour of this argument are recent impressive developments in AI and laboratory robotics.

This improved productivity of science will lead to profound societal benefits, for example better food security or better medicines. I envision a future where the collaboration between human and robot scientists will produce better science than they can by themselves – after all, teams of humans and computers play better chess together than either alone.

GOT A FEW MORE MINUTES?

OECD Conference on Artificial Intelligence – “AI: Intelligent Machines, Smart Policies” -Paris, 26-27 October 2017

ABOUT THE AUTHOR

Ross D. King is Professor of Machine Intelligence at the University of Manchester, UK. His main research interests are in the interface between computer science and science. The research achievement he is best known for of is originating the idea of a ‘Robot Scientist’: using laboratory robotics to physically implement a closed-loop scientific discovery system.