Learning Through Human Feedback | DeepMind


Ever heard of the paperclip problem? It's probably the most boring way ever imagined for humans to become extinct, and DeepMind and OpenAI take it very seriously. Their most recent approach avoids the problem altogether by not specifying a utility function, but instead training the network using human feedback:

...these results demonstrate one method to address this, by allowing humans with no technical experience to teach a reinforcement learning (RL) system - an AI that learns by trial and error - a complex goal. This removes the need for the human to specify a goal for the algorithm in advance. This is an important step because getting the goal even a bit wrong could lead to undesirable or even dangerous behaviour. In some cases, as little as 30 minutes of feedback from a non-expert is enough to train our system, including teaching it entirely new complex behaviours, such as how to make a simulated robot do backflips.

It's a fascinating piece of research, and well-documented. As usual, the example applications are toys, but you can easily imagine armies of humans employed to train algorithms. Commercial applications abound.


Want to receive more content like this in your inbox?