Google DeepMind and OpenAI Team up to Prevent Robot Revolt

Google DeepMind and OpenAI have teamed up on project we can all get behind– preventing an AI rebellion. While artificial intelligence (AI) research is still in its early stages of development (though there is some disagreement surrounding this assessment), humanity has been worrying about the possibility of an apocalyptic robot revolt for decades, which is why the two leading AI initiatives have decided to tag-team and attempt to nip this problem in the bud.

DeepMind, the company behind AlphaGo, and OpenAI, which is partially funded by Elon Musk, have released a research paper detailing a new and potentially safer method of machine learning. The method involves introducing human feedback and preferences into an AI’s deep reinforcement learning system, which should lead to fewer surprises than alternative methods that allow AI to teach itself how to perform tasks. The problem that the team is tackling is that it’s hard to train AI agents to perform complex tasks in predictable and desired ways. For instance, Wired reports that when OpenAI set an AI agent to play a boat racing game, it figured out a way to cheat and score points by driving in circles instead of following the course.

The response they’ve come up with to that issue is to teach AI how to define tasks according to human preferences by training its reward function using human input. The team did so by hiring human contractors to compare short video clips depicting the AI’s behavior and select the most desirable action. After training the AI with thousands of bits of feedback, the agent was able to learn to play simple video games. For example, using less than an hour of human time, the team reported they were able to successfully teach “complex novel behaviors”, like getting a simulated robot named Hopper to perform a backflip.

The downside to this approach is that it is still time-consuming and inefficient, because it requires a significant amount of human oversight. However, its limited successes suggest a way to train AI to align its activity with human preferences. And in any case, gaining expertise in learning how to interact with machines could prove to be very useful in a future where superintelligent machines are a reality. “If you’re worried about bad things happening, the best thing we can do is study the relatively mundane things that go wrong in AI systems today,” OpenAI researcher Dario Amodei said to Wired.

Want a FREE iPhone 7? Click here to enter our monthly contest for a chance!
Follow us on Apple News by pressing the (+) button at the top of our channel

from iDrop News http://ift.tt/2sIPgZZ
via IFTTT