By Super Neuro
Google's Deep Mind AI team has come up with another killer product, training AI to challenge humans' spatial perception abilities. They actually defeated professional human players in a maze game.
It’s terrifying to think about it. In the future, when we play games, we really won’t be able to tell whether the opposing player is a human or AI. It’s best not to teach AI to pretend to be a girl with a female voice to deceive our feelings.
How did Deepmind do it?
Deep Mind, which is familiar with "The Art of War" by Sun Tzu, has been well aware of the principle of knowing both yourself and the enemy since Alpha Go. Therefore, if you want AI to have the same navigation ability as humans, you must first understand how the human brain navigates.
How does the human brain navigate?
Exploring the mysteries of the human brain can be said to be a task with a difficulty coefficient of more than 9.9, but Deep Mind, standing on the shoulders of giants, found the answer to the problem effortlessly.
Special thanks should be given to the British-American scientist John O'Keefe, winner of the 2005 Nobel Prize in Physiology or Medicine, and May-Britt and her husband Edvard Moser from Norway.
Through unremitting efforts and hardships, they found the core of the human brain's GPS - "grid unit", which made it easy for Deep Mind to discover the key to the human brain's spatial perception. Grid unit is a kind of neuronal organization composed of countless grid cells. Its main function is to help humans and most mammals locate and navigate themselves.
When we move, the grid cells will constantly update the current location and surrounding environment, record the walking path and historical location, and then draw a virtual map in the brain to help the brain determine the location and direction.
Every time you go to a new place, the grid cells will automatically draw a new map. That is to say, the more places you go, the more complete the map will be, and the wider your spatial perception range will be.
AI version of grid unit
Inspired by this achievement, the Deep Mind team worked with UCL (University College London) scientists to develop a recursive neural network system. In simulating the process of mice finding directions, the system miraculously discovered that the neural network produced some features similar to those of grid cells when determining positions.
(Note: The grid cells generated by the neural network are very similar to humans)
Through further research and improvement, they successfully created a neural network version of the grid unit (also called AI grid unit). The AI grid unit can not only determine its own position, but also find the best route to the target point in a complex environment.
This discovery made the Deep Mind team overjoyed, and they urgently needed an opportunity to show off their real technology. Therefore, the team decided to skip mice and gorillas and declare war directly on humans.
In a maze game against professional players, the AI equipped with a neural network version of grid units not only found the direction in the complex environment, but also accurately calculated the distance to the target point and found the fastest route (including some other routes that took a longer time), successfully defeating humans.
Although it was only a victory in a virtual environment, it means that AI has the ability to find its way in real scenes without the help of external data such as GPS. Therefore, while Deep Mind has been successful, it is undeniable that this is a milestone victory.
Previously, AI has only been proven to excel in areas such as image recognition and board games, but rarely has it surpassed humans in subjective cognitive areas such as navigation. This time, AI has "evolved" into an automatic navigation function, proving that AI can not only replicate the brain, but also understand it well.
It is foreseeable that AI will have more possibilities in the future. Relying on its powerful computing and learning capabilities, it can come up with several solutions to the same problem and find the best answer.
This is great news for those who are struggling with what to eat for lunch.
Of course, if we continue to train AI to adapt to various gaming environments, coupled with the hard-to-distinguish voice of Google Assistant, it will be even more difficult for us to distinguish whether the game player on the other side is a young lady or AI.
Sad.