It’s a commonly held belief that computers will become sentient and take over the planet, enslaving humans and reigning supreme until a plucky band of unlikely heroes shuts that whole thing down. That almost certainly won’t happen, but that doesn’t make the occasional surprise discovery that machines are getting smarter any less unnerving. Like, for example, when Facebook decided to teach its chatbots to haggle, they wound up inventing a shared language to do it.
Let’s table the fact that Facebook was teaching robots to haggle over the price of stuff using and focus on the machine learning aspect.Machine learning is simple — the computer learns by doing, just like the rest of us. Either it gets it right, and gets a pat on the head, or gets it wrong, is told that we’re not mad, just disappointed, and gets sent back to try again. Basically, you make a computer do something over and over again, eventually it learns to do it the right way. The study set up two chatbots to negotiate until they came to a satisfactory conclusion, and to do it in natural language, which yes, is a bit creepy. Much creepier though is this line on the first page:
To avoid diverging from human language, we interleave reinforcement learning updates with supervised updates.
Translation: “Our robots started talking to each other in a language we didn’t understand, so we regularly had to force them to stop doing that.” The researchers go into a little more detail later in the paper:
…we found that updating the parameters of both agents led to divergence from human language.
In a way, this isn’t surprising. Computers aren’t humans; they speak in ones and zeroes, not via sound. If they’re talking to each other, and reinforcing each other, then they’re inevitably going to create a more efficient way of communicating. Sadly there’s no indication of what this language actually was, but we suspect this won’t be the last time researchers discover their programs have, in their own way, a mind of their own.
(Via The Atlantic)