What happens when we sleep? While the body rests, the brain continues to work. This is what Google wants, the tech giant is granting computers the ability to dream when it sleeps.
In humans, sleep helps consolidate memory, shifting out short-term memory into long-term storage. The process goes through three distinct phases: stabilization, enhancement and integration. With the three, we can transfer memories into what can be easily recalled in the future.
These distinct phases our brain can store information like a file cabinet. This is what Google is doing with its DeepMind AI. The goal is to make robots to have the same ability.
For all this time, we've created AI to learn from supervised methods. This is where we make AI to analyze data to look for patterns. Supervised is a fairly straightforward approach in teaching machines but apparently not how humans learn. Humans use an approach that is called "unsupervised learning" in which the we experiment things on our own to determine how different courses of action affect our goals.
Google's attempt to make computers dream, follows the wake of neuroscientific discoveries that revealed the importance of dreaming for memory consolidation.
So what the attempt does, is to make AI dream to improve their rate of learning.
Google's DeepMind first major successes started from video games like Breakout and Asteroids. Those classic games have taught the AI to understand that in order to beat the game, it needs to know more than just sequences. This time, Google makes use of this computer achievement by bringing it to the next level.
Researchers at DeepMind are primarily concerned with unsupervised learning because it holds the best hope for creating AI with human-like general intelligence.
Read more: When Machine Learns, It All Started With A Human's Dream And A Good Imagination
While DeepMind is already an advanced piece of engineering able to play games, it still won't be able to beat any humans in complex games. To make AI smarter, DeepMind is making AI to dream about games it can already master (consists primarily of scenes from Atari Video games), to then having it rehearse the tasks that are separated into sections to visualize the path to victory, and to repeat the process all over again.
The goal here is to make AI learns as humans do, which is through experimentation and repetition. This is where AI is made to shift from supervised learning to unsupervised learning where it needs to experiment by doing analysis on different courses of actions.
This type of learning is far more complex and time consuming due to the number of variables, which are infinite. However, it can be an ideal solution to make AI learn, when it is inactive, or sleeping. Thus making it dream.
While the attempt is still in development and experimental, so far, researchers have reported a 10x speed increase in the rate of learning over supervised training.
Dreams play an important role in learning and memory formation. The end game for making AI dream, is not to make it able to dream like us humans. The challenge here is to make AI able to learn from the real-world situations that are more challenging and unpredictable. When it's finally able to learn by itself unsupervised, AI can adapt to a given situation all by itself.
Further reading: Google and Facebook: When Computers Dream, Humans Are Fooled