We all dream. Luckily, we do not always know what other people dream, and our box of dreams is conveniently closed to the peeping of the people around us. Frankly, neuroscience discovered that our dreams are pretty much a dark and doom affair anyway. But DeepMind, Google’s powerfull AI (artificial Intelligence) dreams… and wakes up a lot wiser.
Reply awkward situations
Scientists and psychologists discovered that dreams, from a purely neuroscientific perspective were mostly negative or threatening. That awkward dream of you being lost in a maze, of turning up way late at work completely naked, or incapable of outrunning that fire spitting beast are more rule than exception. The latest neuroscience theories claim that dreams strengthen the neuronal traces of recent events. Negative feelings help cement memories deeper into the brain. This enhances memory formation. Long story short: your brain packages learnings, plays them back in your unconsciousness, and links the learning to a negative emotion (the bad dream) so the learning gets filed more properly in your memory. Neat.
Play it again, Sam
Often we get the same dream over and over again, with slight variations in the storytelling. Our brain is trying to find a solution for a complex problem. Neuroscientists are convinced that this helps you finding the solution quicker. Quick replay, fast forward. Add variation. Repeat. Learn, store. Dream.
Contrary to other AI’s, that are getting fed controlled pathways to the solutions, Googles Deep Mind is programmed to learn by itself. It detects and plots its own way to a solution, learning to solve complex puzzles all by itself. DeepMind experiments on its own to detect how different variations of action within a situation affects the outcome. That is how it learns. Unsupervised learning eats up time, because it involves variations and experimentation. However, the learning method allows a lot of future time gain: the robot learns how to learn, and becomes able to learn by itself. Self-learning is a requirement in the stages to true intelligence.
Dreaming is parallel processing
DeepMind uses a dream like state to highlight certain extremely challenging parts of its deep learning, and repeat them endlessly, adding slight variations. Rather than solving the whole puzzle over and over again, it will replay peculiar parts, and variate, much like humans do while they dream. After a certain number of loops, the challenge is solved, the method learned, the result stored for future reference: expertise is achieved. The researchers at DeepMind were able to boost DeepMind’s speed of learning with a whopping 10x speed.
Better Robots through better understanding humans
So, adding a dream feature enables AI’s to function better and quicker. Dreamlike loops are one of the first deeplearning on how to learn that machines learn to master. It reminds me of my good friend John C. Havens who in all his wisdom says that “to build better machines, we will have to better understand humans”. My question is: who will understand who the quickest?
Sssst…. your robot dreams, probably about you!