How Curiosity (Sometimes) Killed The A.I.

The appeal of Artificial Intelligence is its ability to learn. Feed it a data set for which there are (known or unknown) solutions, and watch it learn! A lot of this, however, depends on humans feeding the AI the right data sources, such as translations of the same documents in different languages.

What if A.I. could want to learn on its own, though? What if it could be given that most human (and, allegedly, feline) trait called curiosity?

New research by non-profit lab OpenAI set out to answer this question. In order to simulate curiosity, they encoded a reward system into the Machine Learning Algorithms to reward the A.I. for encountering new stimuli. More specifically, the A.I. would attempt to guess what the next frame of a video game or television program would look like, and be rewarded for how wrong it was (under the reasonable assumption that a very wrong guess means it has encountered something new, something it did not see before that moment, such as a new level of a game, or new channel on television).

Driven by this, the A.I. played over 50 video games and watched a lot of television (controlled by means of a virtual remote control). It even managed to beat some of these games.

On the downside, the A.I. also sometimes deliberately died in a game, in order to see an end game screen it had not seen before, and became the worst sort of Couch Potato for television, flipping endlessly through channel lineups just to keep seeing new things. In short, it developed something roughly akin to FOMO, the Fear of Missing Out, in this case missing out on new stimuli.

Clearly, then, pure curiosity will likely not be an ideal approach for Machine Learning. However, combined with more classical learning approaches, this may give A.I.'s a taste of one of the very things that makes our own learning process unique: to feel rewarded for the mere act of learning new things and experiencing new things in and of itself.

#AI #ArtificialIntelligence #MachineLearning