Nocturnal sleep or daytime nap helps human to fasten memory, emigrating short-term memory to long-term storage. Your brain shifts memories into an organized filing system for easier recall in the future through these distinct processes- stabilization, enhancement, and integration. Inspired by this neuroscientific discovery, Google’s DeepMind is developing a new technology that lets robots dream in order to improve their rate of learning.
Primarily, the element of these AI dreams consists of scenes from Atari video games. DeepMind’s first major success came in the form of classic video games, such as Breakout and Asteroids. Those games taught artificial intelligence the further simple sequences necessary to beat the game, but also set the base for the supervised learning system used now.
Given that AI can already beat humans in games like Chess or Go, DeepMind is still not up to the task of dominating humans at complex games like Labyrinth or Starcraft. AI dream consists of repeating entire sections of these games to visualize the path to victory and repeat until it becomes an expert, where our dream consists of threatening issues and embarrassing situations.
The goal is simple, moving artificial intelligence toward learning as human do. To transfer from supervised learning to unsupervised learning, it’s important to teach robots to experiment and explore how different courses of action affect the upshot. Due to an infinite number of variables, this type of learning is way more time consuming, suggesting it as an ideal solution for periods of inertia, or dreaming.
In the AI space, making a robot dream is still a thriving area of research. However, scientists so far report a fascinating 10x speed increase in the rate of learning over supervised training.
This is experimental at this point. However, given the social role that the researchers continue to imagine for robots, it does seem that robots could soon dream of showing up for work naked, which would actually be beneficial for the future of AI.