Scientists for DeepMind, the AI project owned by Google parent company Alphabet, seem to have run into some roadblocks recently regarding its projects development. According to a piece written by Gary Marcus for Wired, “DeepMind’s Losses and the Future of Artificial Intelligence,” DeepMind lost $572 million last year for its deep pocketed parent company and has accrued over a billion dollars in debt. While those kinds of figures are enough to make the average parent feel much better about their child’s education dollars, the folks at Alphabet are starting to wonder if researchers are taking the right approach to DeepMind’s education.
So what’s the problem with DeepMind? Well, for one thing, news of DeepMind’s jaw-dropping video game achievements have been greatly exaggerated. For instance, in StarCraft it can kick ass when trained to play on a single map with a single character. But according to Marcus, “To switch characters, you need to retrain the system from scratch.” That doesn’t sound promising when you’re trying to develop artificial general intelligence. Also, to learn it needs to acquire huge amounts of data, requiring it to play a game millions of times before mastery, far in excess of what a human would require. Additionally, according to Marcus, the energy it required to learn to play Go was similar “to the energy consumed by 12,760 human brains running continuously for three days without sleep.” That’s a lot of human brains, presumably fueled by pizza and methamphetamine if they’re powered on for three days without sleep.
A lot of DeepMind’s difficulties stem from the way it learns. Deep reinforcement learning involves recognizing patterns and being rewarded for success. It works well for learning how to play specific video games. Throw a little wrinkle at it, however, and performance breaks down. Marcus writes: “In some ways, deep reinforcement learning is a kind of turbocharged memorization; systems that use it are capable of awesome feats, but they have only a shallow understanding of what they are doing. As a consequence, current systems lack flexibility, and thus are unable to compensate if the world changes, sometimes even in tiny ways.”
All of this has led researchers to question whether deep reinforcement learning is the correct approach to developing AI general intelligence. “We are discovering that the world is a really fucking complex place,” says Yuri Testicov, DeepMind’s Assistant Director of Senior Applications. “I mean, it’s one thing to sit in a lab and become really great at a handful of video games, it’s totally another to try to diagnose medical problems or discover clean energy solutions.”
Testicov and his fellow researchers are discovering that the solution to DeepMind’s woes may not come from a new approach to learning, but instead, the public may need to lower the bar on expectations. “We’re calling on the people of earth to simplify and dumb down,” adds Testicov. “Instead of expecting DeepMind to come along and grab the world by the tail, maybe we just need to make the world a little easier for it to understand. I mean, you try going to the supermarket and buying a bag of tortilla chips. Not the restaurant kind but the round ones. Not the regular but the lime. Make sure they’re low sodium and don’t get the blue corn. That requires a lot of complex awareness and decision making. So, instead of expecting perfection, if we send a robot to the supermarket and it comes back with something we can eat, we say we’re cool with that.”
Testicov has some additional advice for managers thinking about incorporating AI into the workplace. “If you’re an employer and you’re looking to bring AI on board, don’t be afraid to make accommodations for it, try not to be overly critical of job performance, and make sure you reward good work through positive feedback and praise,” says Testicov. “Oh sorry, that’s our protocol for managing millennials. Never mind.”