Your Cart

‘Landmark Event’ In Artificial Intelligence: DeepMind Trains Itself

Posted by Colin Clark on

Go Master Ke Jie defeated by AlphaGo

WASHINGTON: When a computer system defeated one of the greatest masters of the complex game known as Go last year, the world gasped. Experts had said months before that such an event would not occur in their lifetimes.

Last night, the magazine Nature published an article by DeepMind, the Google company behind that breakthrough, claiming that its AlphaGo Zero computer program taught itself in three days what had required its predecessor, AlphaGo, over half a year, absorbing data from more than 100,000 human amateur and professional games.

“Over the course of millions of AlphaGo vs AlphaGo games, the system progressively learned the game of Go from scratch, accumulating thousands of years of human knowledge during a period of just a few days. AlphaGo Zero also discovered new knowledge, developing unconventional strategies and creative new moves that echoed and surpassed the novel techniques it played in the games against Lee Sedol and Ke Jie,” according to an accompanying paper by those who helped develop AlphaGo Zero.

At an event for its new Center for Artificial Intelligence and Autonomy, CNA’s principal research scientist, Andy Ilachinski, cited the DeepMind work as “one of the landmark events in the century we are now in.” How big will its effects be? “I will simply say, in my opinion, AI will have a greater impact in our lifetime, certainly our children’s lifetime, on human culture than anything in our history.”

While Artificial Intelligence has been a key part of the Pentagon’s Third Offset Strategy, much of the discussion about it has been theoretical. It would greatly aid our ability to assess data and make decisions much more rapidly. In fields such as cyber warfare and missile defense, AIs could provide unmatchable overmatch by combing through, analyzing and recommending actions in seconds instead of minutes and hours.

Google image, screencapped from Youtube

The original AlphaGo’s win-loss record against the human champion.

Not only did AlphaGo Zero teach itself, it obliterated its older sibling, winning 100 games in a row. Poor AlphaGo didn’t win a single game. Imagine how human Go masters will fare against this latest AI incarnation.

Defense policymakers and the wider public will have to grapple with the fabulously fast –and sure to increase in velocity — pace of AI developments, Ilachinski said. If we are to deploy these AI systems we must change how we buy weapons and how we manage them. “You can’t precertify where these systems are going to go,” he said, so DoD and our society as a whole will have to change their expectations of how quickly they must act and build the tremendous speed and power of the AI’s work into our planning.

Coping with such concepts while the field is still so new is challenging for the Pentagon, said Shawn Steene, senior force planner for emerging technologies in the Office of Secretary of Defense.

“I would argue to some degree we’re limited somewhat by the physical reality and our own creativity in the application of these capabilities,” Steene said.

Will the vaunted ability of the US military. to change its organization and doctrine to cope with change keep pace with the speed and power of Autonomous Intelligence? No one at today’s event was sure, but Joseph Horvath, who works on unmanned systems in the Navy’s Warfare Integration Division, pointed to the trust and freedom given to soldiers, sailors, airmen and Marines. “We believe in our troops, we believe in their skills, their training, their experience, their ingenuity — all a function of our democracy and our way of life,” he said. “We see that teaming (with AI) as what will give us the advantage.”

What Others Are Reading Right Now