When a robot goes to sleep, what does it dream about? Atari games, apparently.

In humans, nocturnal sleep (or even daytime naps) helps to consolidate memory, alive concise memory to abiding storage. Through three audible processes — stabilization, enhancement, and affiliation — your brain transfers memories into an uber-organized filing system for easier recall in the future.

First, stabilization encodes a memory in all of six milliseconds. Next, the brain enhances the memory over minutes, hours, or a full day to coalesce it into your longer-term memory. Finally, the affiliation action occurs, which is where the brain adds new pieces of memories into the old ones — acting as a sort of archival system, not unlike a file cabinet.

The end game is to get robots to do the same.

Google’s first major successes with DeepMind came in the form of archetypal video games. Games like Breakout and Asteroids taught the AI the rather simple sequences appropriate for assault the game, but also laid the foundation for the supervised acquirements techniques used today.

In humans you have to crawl before you walk; with robots, you have to beat Asteroids before you beat facial apprehension or cancer research.


As avant-garde as DeepMind is, it’s still not up to the task of assault humans at more circuitous games, like Starcraft or Labyrinth. Where we dream of awkward situations and aggressive issues, AI dreams abide of call entire sections of these games to anticipate the path to achievement and repeat it until accomplishing expertise.

The goal is to move bogus intelligence toward acquirements as humans do: through experimentation. The shift from supervised acquirements — where AI analyzes supplied data, attractive for patterns — to unsupervised acquirements involves teaching robots to agreement and assay how altered courses of action affect the outcome. This type of acquirements is far more time arresting due to an absolute number of variables, which makes it an ideal band-aid for periods of inactivity, or dreaming.

It’s still a beginning area of analysis in the bogus intelligence space. But so far advisers report an absorbing 10x speed access in the rate of acquirements over supervised training.

Although beginning at this point, the early after-effects seem to advance robots assuming up for work naked, at least in dreams, could be benign to the future of bogus intelligence. Or, maybe we all just watch too much Westworld.