“Tickling Your Funny Bone: A Deep Learning Substitute Aids AI Agents in Game-Playing the Real World”
“A Deep Learning Alternative Can Help AI Agents Gameplay the Real World”
“Most of the big leaps in AI have a dirty secret: They’re dependent on human labeling. In order for an algorithm to learn that a photo depicts a cat, someone first has to point out the cat.”
That’s right, folks! The so-called advanced AI we all admire and fear, the technology that’s supposed to take over the world, learns just like a toddler stumbling around a living room. Just like how a tot needs an eager parent to point at the chubby, lazy furball on the mat and say, “That’s a cat!”, our sophisticated algorithms need similar hand-holding. How’s that for a billion-dollar industry?
As stated in the aforementioned WIRED article, this desperate need for human supervision has been a major stumbling block on the road towards self-sufficient AI. But don’t despair just yet, there’s a glimmer of hope. DeepMind, the brainy folks behind AlphaGo, have whipped up a potential solution in the form of ‘reinforcement learning from pixels’ or RLØ, as they like to call it.
This “unassisted” learning method equips AI to understand the world via unlabelled images — kind of like how we humans learned most things when we were little, but without all the crying and diaper changes. RLØ is a unique agent that “teaches itself to understand and interact with a 3D environment, similar to how a human first learns about the world in their early years.”
Showing off its skills in a virtual environment designed to mimic real-world physics, RLØ demonstrated that it could teach itself to play hide-and-seek, just by exploring its surroundings. Quite impressive, especially when you recall the view from up close and personal as your grandmother looked for her glasses, which were on top of her head the entire time.
But before we get too excited, let’s keep our feet on the ground and hats on our heads. Despite the breakthrough, we’re still a long way from getting an AI that can completely understand and function in the real world. Compared to our brains, AI is still in its learning socks, while we’ve graduated to enlightenment loafers, so to speak. But hey, watching AI toddle and tumble along the way sure keeps things interesting.
So, here we are dear readers — on the brink of technological advancement, yet chained to the very human necessity of labeling to keep this AI beast fed. And even though RLØ is miraculously teaching itself the ropes (quite literally, in the hide-and-seek challenges), it’s still us calling the shots and setting the game rules. We might not be as obsolete as we thought…yet.