How to build a better, safer artificial intelligence

Some of today’s smartest people are scared of smart machines.

In a BBC interview, experimental physicist Stephen Hawking said: “I think the development of full artificial intelligence could spell the end of the human race.” Entrepreneur and visionary Elon Musk was more pointed. “With artificial intelligence, we are summoning the demon,” he said.

Earlier this month, Musk announced the founding of a non-profit AI research company, called Open AI, to “advance digital intelligence in the way that is most likely to benefit humanity as a whole,” according to the company’s mission statement.

A few weeks earlier IBM began promoting SystemML (machine learning), a proprietary machine-learning technology under an open-source license, available through the Apache Software Foundation. SystemML is built to recognize patterns in Big Data.  It is designed to work with Spark software, which helps process data as they arrive from multiple sources, such as fitness trackers or, ultimately, EMR systems.

Both efforts are aimed at making AI safer.  Louis Rosenberg may have the safest idea of all.   Rather than create an intelligence that “will not share our interests and would likely be as foreign from us as an alien intelligence,” Rosenberg suggests amplifying human intelligence with an algorithm that turns groups of people into super experts.

Rosenberg calls this approach swarm AI.  Rather than taking people out of the decision making loop, it makes them a critical part of it.  Take the humans out. And the AI. Stops.

Rosenberg’s algorithm, a product of his company called Unanimous AI, brings people together. It leverages their intuitions, emotions, sensibilities – and above all, knowledge – to draw conclusions that are better than any one expert. 

This approach may be especially suited to medicine and health IT, which could be among the first to feel  the effects of AI.  Learning machines have been proposed for examining medical records, sharing data among information systems, and drawing insights that can help physicians make better decisions in the diagnosis and treatment of patients. IBM is grooming its Watson Health to help physicians make diagnoses.  A San Francisco startup called Enlitic is implementing a deep learning algorithm in Australian and Asian imaging clinics to help physicians spot patterns of disease in medical images.

These systems are based on conventional AI, which depends on machine learning.  Rosenberg is promoting swarm artificial intelligence as an alternative, one that has all the advantages but none of the drawbacks of conventional AI. It may, in fact, be the only kind of artificial intelligence that has a prayer of being approved by the FDA in the foreseeable future .

 “I think the people who are really excited about conventional AI systems underestimate what doctors are really doing (when they make diagnoses). There is more to it than just simple rule-based decisions,” Rosenberg says. “With swarm AI, we are trying to have the benefits of amplified intelligence while keeping human sensibilities and human values and human interests deeply ingrained in the system as opposed to just replacing them.”

Early tests of swarm AI for medical diagnosis have been promising.  In one study, a collective intelligence of radiologists proved superior when interpreting mammograms, reducing false positives and false negatives. This swarm AI overcame “one of the fundamental limitations to decision accuracy that individual radiologists face,” the authors concluded in a peer-reviewed paper published by the Public Library of Science. Their study demonstrated the collective or swarm intelligence could improve mammography screening and has the potential to improve many other types of medical decision-making, “including many areas of diagnostic imaging.”

In another study, a dozen radiologists increased their ability to correctly diagnose skeletal abnormalities.  The researchers reported at the ninth international conference on swarm intelligence that the “algorithm’s accuracy in distinguishing normal vs. abnormal patients was significantly higher than the radiologists’ mean accuracy.”

You can argue that the real world doesn’t have the luxury of bringing groups of radiologists together to develop a consensus on every case.  But swarm AI wouldn’t be needed for every case, Rosenberg says.

Routine cases could be handled by individual physicians who, when stumped by a complex case, could tag it for later examination by swarm AI.  This would improve diagnosis, while empowering team members.

“These swarms would help the group come to a decision, as opposed to just taking a vote where the vote might just reveal the differences in the group,” Rosenberg says. 

To promote swarm AI, Rosenberg last year founded Unanimous AI, a company created with the goal of enabling groups to “think together”. With this application of swarm AI, Rosenberg borrows a page from nature’s playbook wherein species accomplish more by participating in flocks, schools, colonies and swarms than they can individually.  Unanimous AI offers the unique infrastructure, he says, by which people can form intelligent swarms. 

 “We know that groups are smarter than individuals and we also know that nature has addressed this by having swarms (of animals) make optimal decisions,” he says.  “We are just connecting people in a way that can harness their diverse perspectives and opinions and knowledge.”

Swarm AI is distinct from voting in that once members of a group have cast their votes, the process is over.  Rather than polling members, swarm AI uses a cyber “puck” that members of the group constantly pull or push. The puck moves in real-time on the computer screens viewed by the group toward or away from different possibilities. It does so depending on the beliefs and conviction of group members about the correctness, for example, of a diagnosis.

“They are all at the same time watching the members of the group express their conviction, where the group converges on the decision that optimizes their collective will,” Rosenberg says.

The potential for improved analyses without the risks of machine intelligence make swarm AI an attractive alternative to machine learning. Now in beta testing, swarm AI could emerge commercially as early as next year.