Google DeepMind is planning to make their artificial intelligence agents "dream" in order to learn faster just like humans do when they sleep.
The search engine giant's AI company wants their machine learning technology to learn even faster through the power of dreaming. When humans sleep, their brains can consolidate more memory and shift what they learned during the day to their long-term storage.
Stabilization happens first during the first six milliseconds and then enhancement happens from several minutes to a few hours. It could also happen for a whole day depending on the kind of information and quality of sleep.
Integration then happens after solidification of the memory is complete. Pieces of new memory are added to the old ones to reinforce and learn faster.
DeepMind also wants the same processes to happen with their AI project. Current machine learning techniques involve supervised feeding of data to the AI. This means that they are fed the data that is needed instead of having them find the data and learn on their own.
Unsupervised learning is possible but it is time consuming considering that there are infinite possibilities and combinations involved. Microsoft is already using "Minecraft" to test their AI agents as they learn to proceed through the levels without dying.
While DeepMind's AI does not technically "dream," their process of conducting periods of inactivity have proven to be effective. Google's DeepMind project with dreaming reportedly resulted in 10 times faster learning for their AI but through supervised training, The Next Web has learned.
DeepMind already tests their AI by having them attempt to play old-school video games such as "Asteroids" and "Breakout." While they have little to no difficulty in winning the said games, they are still not prepared to take on more modern and complex games such as "DOTA 2" or "League of Legends" on their own without any pre-programming.
Dreaming would help the AI learn how to adapt to more challenging situations. Most human dreams are either negative or threatening to the subject as this helps them to learn what to avoid when they are finally awake.
DeepMind also aims to let their AI experience the tough situations during their periods of inactivity. Scientists and researchers theorized that such nightmares in humans help memory to be more solidified inside the brain's storage.
AI that undergo dreaming will have to repeatedly play out levels or complicated sections in a game in order to perfect their skills. It is still supervised learning which means that they are still fed the data they need. In the game's case, they are left to dream in a specific part of a level instead of having to run the whole set which is time consuming.
Google's DeepMind is still aiming for unsupervised learning that will be effective and more efficient compared to the traditional methods today, Extreme Tech reported. Unsupervised learning is believed to be the way to achieve human-like intelligence in the future.
Humans do still undergo some form of supervised learning but the generally experiment with the given information. For instance, one student may learn basic math but another may already start to develop their own calculating techniques through experimentation.
DeepMind said in their findings: "We have shown how augmenting a deep reinforcement learning agent with auxiliary control and reward prediction tasks can drastically improve both data efﬁciency and robustness to hyperparameter settings. Most notably, our proposed UNREAL architecture more than doubled the previous state of-the-art results on the challenging set of 3D Labyrinth levels, bringing the average scores to over 87% of human scores. The same UNREAL architecture also signiﬁcantly improved both the learning speed and the robustness of A3C over 57 Atari games."
The 10 times faster learning speed will certainly help more in speeding up research for Google's DeepMind AI projects. Possible applications could be AI assistants learning while they "dream" or during their inactive times when the smartphones are idle. Although, it still remains to be seen how much power it will require.
Video games could also have more advanced AI that learns the players' moves instead of just having to rely on a pre-set difficulty. It would also help the player improve even further as the AI enemy also improves over time.