“Memories, you’re talking about memories!,” exclaims a dubious Rick Deckard in the 1982 movie Blade Runner. He is hit with a grand epiphany as Dr. Eldon Tyrell drones on about how implanting non-existent experiences in the minds of test subjects gave him much more control on them than previously deliberated. This is how an average day at DeepMind Technologies Laboratory in London sounds like.
Is winning a memory too sweet?
DeepMind Technologies is an artificial intelligence company that was established in 2010. Acquired by Google in 2014 for an acquisition price of $500 million, this startup focuses on getting machines to think and learn. The researchers at DeepMind lets their AI (Artificial Intelligence) play Atari games. The AI isn’t taught how to play games, but it learns the basics of the game after it loses a few times. After many unsuccessful trials, the AI gets better at it as it manages to beat the game with controlled precision and without seeming to break a sweat even as the difficulty is increased drastically. Interestingly, the researchers at DeepMind noticed that rather than learning from its experiences of playing the previous levels, the AI ‘believes’ each level is a new game, training itself to amass knowledge about the physics and environment of the game to get through it. The AI simply had no ‘memory’ of it finishing Space Invaders when it was made to play again.
That’s why the researchers at DeepMind developed the Elastic Weight Consolidation (EWC) algorithm which remembers the key moments in the game that enabled the AI to win and later these experiences are applied when it plays the next level, often adapting and evolving its experiences to beat successive levels of the game. The EWC algorithm helps the AI anticipate as a veteran gamer would. DeepMind AI, with a few lines of code, has succeeded in mimicking the human brain, and perhaps even supersede it with an unexpected rate of learning. Does that make them formidable?
To compromise? Or to attack?
The answer to this question which intrigued thinkers and futurists since ages can be disputed once we consider an experiment conducted by DeepMind. Two AI agents were made to play a game, ‘Gathering’ where they had to collect fruits from a central supply. The agent with the most fruits win. The rules were simple but came with a twist, the agents were also armed with a laser that when used against an opponent, puts the latter out of the game briefly, allowing the aggressive agent to collect more apples. What followed was a game of hostility that revealed the AI’s selfish side as the agents went ballistic on each other, trying to emerge as the winner.
In another game ‘Wolfpack’, two AI agents had to capture a third agent and points were awarded not only to the capturer but also to the other agent that aided in cornering the third. In this scenario where cooperation is rewarded, the agents were observed to conspire together to corner the third. They acted in unison as the game rewarded cooperation rather than aggressiveness, which was the case in Gathering. In the near future when AI will pervade every aspect of our lives, will they strategize against us because the ‘reward clauses’ were miscoded or tampered with? Or would they disregard us because they become too self-aware of their abilities?
The fall or rise of men?
Speaking about abilities, perhaps we have all heard how Garry Kasparov was defeated in chess by an IBM supercomputer that went by the name Deep Blue. Why was this a momentous feat? There are 10^120 possible moves in chess and teaching a machine to play such a game of the seemingly infinite number of possible moves was considered impractical, much less than a machine defeating the then World Chess Champion.
After Kasparov’s defeat in 1997, the next step for AI scientists was to tutor AI the moves of ‘Go’, an ancient Chinese board game which has more possible combination of moves than the entire atoms in the universe. DeepMind’s algorithm AlphaGo relied on reinforced learning to teach itself the complex game and in 2015, it made headlines when it defeated Lee Sedol who is considered one of the best minds at Go, without suffering a handicap. A machine-learned a game from scratch and bested humans at it. But there’s more to it, Go is a much more intuitive game whereas chess is a game purely based on logic. DeepMind’s AlphaGo has demonstrated that AIs are much more than a few lines of code. Apparently, they can develop intuition and perhaps in a few years, even emotions.