‘The hardest puzzle I can think of’
NEUROBIOLOGIST R. CLAY REID ON REVERSE-ENGINEERING THE BRAIN
Koch is more hesitant. “Runaway machine intelligence is something we need to think about more,” Koch, president and chief science officer of the Allen Institute for Brain Science, said. “Clearly, we can’t say let’s not develop any more AI. That’s never going to happen. But we need to figure out what are the imagined dangers and what are the real ones and how to minimize them.”
Allen’s vision is creating an AI machine that would be like a smart assistant, rather than an independent being, “answering questions and clarifying things for you and so forth.” But he admits he has wondered whether it will one day be possible for that assistant or its descendants to evolve into something more.
“It’s a very deep question,” Allen said. “Nobody really knows what it would take to create something that is self-aware or has a personality. I guess I could imagine a day when perhaps, if we can understand how it works in the human brain, which is unbelievably complicated, it could be possible. But that is a long, long ways away.”
Human brains
Allen’s brain philanthropy took shape when his mother, Faye, a former elementary school teacher, became ill with Alzheimer’s. Allen was very close to her and was devastated when she began to regularly exhibit symptoms in 2003.
Within months, he had founded the Allen Institute for Brain Science and gave it the mission: Figure out “how information is coded in the brain.” Allen, who has committed a total of nearly $500 million to the institute since then, thought that gathering great minds under one roof, all focused on the same goal, could accelerate the process of discovery.
Allen’s “big science” strategy has attracted and significantly increased the salaries of some of the world’s top talent — including R. Clay Reid, a neurobiologist who left Harvard Medical School in 2012 to continue his work on how vision works in the brain. “The brain is the hardest puzzle I can think of, and never before has such a large group been directed to reverse-engineer how it works,” he said.
The Allen Institute also has pioneered a number of other approaches uncommon in biology research.
First, the brain institute started with data, not a hypothesis. Not just ordinary big data but exabytes of it — billions of gigabytes, the scale of global Internet traffic in a month — detailing the growth, white matter and connections of every gene expressed in the brain. Researchers spent their first few years painstakingly slicing donor brains into thousands of microthin anatomical cross sections that were then analyzed and mapped.
By 2006, the institute’s scientists had created the most comprehensive three-dimensional map of how the mouse brain is wired and released that atlas to the public, as promised. By 2010, they had mapped the human brain.
Now many of the institute’s 265 employees are turning to more tangible problems, studying autism, schizophrenia, traumatic brain injury and glioblastoma, a rare but particularly aggressive type of brain tumour, as well as projects to understand the nature of vision.
Artificial brains
All along, Allen has been backing parallel projects in artificial brains. He wondered whether it might be possible to encode books — especially textbooks — into a computer brain to create a foundation upon which a machine could be a digital Aristotle, using a higher level of knowledge to interact with humans.
“I wasn’t aiming to solve the mystery of human consciousness,” he explained in his 2011 memoir. “I simply wanted to advance the field of artificial intelligence so that computers could do what they do best (organize and analyze information) to help people do what they do best, those inspired leaps of intuition that fuel original ideas and breakthroughs.”
That idea grew into the Allen Institute for Artificial Intelligence (or AI2 as it is called by its employees), which opened its doors on Jan. 1, 2014, and currently has 43 employees — a number of them recruited from places like Google and Amazon. Allen hasn’t publicly announced the exact amount of his investment, but Etzioni said it is in the tens of millions of dollars and is growing.
Over the past year, Etzioni and his team have created Aristo. The institute’s first digital entity now is being trained to pass the New York State Regents high school biology exam.
Not only do the engineers have to figure out how to represent memory, but they have to give this entity the ability to parse natural language and make complex inferences. It’s not as easy as it sounds.
“It’s paradoxical that things that are hard for people are easy for the computer, and things that are hard for the computer any child can understand,” Etzioni said. For example, he said, computers have a difficult time understanding simple sentences such as “People breathe air.” A computer might wonder: Does this apply to dead people? What about people holding their breath? All the time? Is air one thing? Is it made up of a single molecule? And so on. The data that Aristo possesses doesn’t add up to the wisdom an elementary school child has accumulated about breathing.
Another test question would require an AI program to interpret this narrative: “The ball crashed through the table. It was made of Styrofoam.” A human might grumble about pronoun-antecedent ambiguity but still quickly conclude that the second sentence described the table. Now if the second sentence were changed to “It was made of steel,” the human would conclude it described the ball. But that type of logic requires a large amount of “common sense” background knowledge — about materials like Styrofoam, steel and wood and how they work, furniture, how balls roll and so forth — which has to be explicitly taught to computers.
So far, Aristo has passed the first-, second- and third-grade biology tests and is working his way through the fourth. The last time Aristo took this test, a few months ago, the grade was about a C. Or, more precisely, 73.5 per cent.
Etzioni says that’s pretty good — for a computer. Sounding like a glowing parent, he said, “We’re very proud he has started to make measurable progress.”
But he estimates that Aristo needs five more years to pass the eighth-grade test.
After that, who knows?
Convergence
The artificial intelligence researchers and their counterparts in brain science are in a kind of race, Allen says, and their work one day will converge — although to what end he’s not sure.
Eric Horvitz, director of Microsoft Research’s main lab in Redmond, Wash., and a past president of the Association for the Advancement of Artificial Intelligence, announced in December he would fund a major research project on the potential effects of AI on society.
Led by Stanford University historians, the study would run for 100 years. The first report is scheduled to be completed in 2015 and subsequent ones will be published every five years, updating technological progress and recommending guidelines.
“If we could design them from the ground up to be supporters of their creators, they could become very strong advocates of human beings and work on their behalf,” said Horvitz, who sits on the board of AI2.
But could those beings ever become self-aware?
Koch, the expert on the subject, isn’t sure.
On the one hand, he believes consciousness is a property of natural systems: “The job of the stomach is digestion, the heart to pump blood. Is the job of the brain consciousness?”
“In principle, once I replicate this piece of highly organized matter I should be able to get all the properties associated with it,” he said. But he said scientists and philosophers aren’t in agreement about what is the right way to do this, under what circumstances and whether it should be done at all.
Two iconic works of science fiction of the 1950s address that question in an ominous way. In Isaac Asimov’s “The Last Question,” humans ask a supercomputer how to save the world until they are gone. Only the machine is left when it comes up with the answer and in the end it commands, “Let there be light ...” In Fredric Brown’s “Answer,” a “supercalculator” made up of all the machines on 96 billion planets is asked: “Is there a God?” Its answer: “Yes, now there is a God.”
“I don’t think we’re building a god by any means,” Etzioni said. “We’re building something on science. The computer is an assistant — not someone you ask, ‘Solve cancer and get back to me.’
“I think it’s going to be something very sophisticated with vast amounts of information, but I still think of it very much as a tool.”