Today we increasingly hear about certain successesvarious artificial intelligence systems. And, despite the fact that developed neural networks are getting better and smarter every day, they are still very far from work similar to the functioning of our brain. And the point here is not even that we don’t have enough hardware to simulate the brain. It’s just that AI does not work at all like our consciousness. And in order to achieve this, it is necessary to revise the approach to the development of machine neural networks. To start, create an AI that will work just like a baby’s brain.
Why AI is not as “smart” as we think
Almost all AI systems that we know abouttoday, simulated based on traditional computer algorithms. They see the world through the prism of binary code - zeros and ones. This is great for complex computing, but according to Professor Alice Parker, who has been involved in AI developments for more than 10 years, humanity is quickly approaching the limit of computing capabilities.
Since the first invention of AI and developmentdeep learning further improvement of such systems went very slowly. In order to fully reveal its potential, AI should not only work faster, it should independently respond to events and learn in real time. And for this to happen, we need to reconsider our approach to the design of artificial intelligence systems.
How to make AI smarter
To solve the indicated problem, professorParker and her colleagues use the most advanced learning system nature has ever created: the human brain. And the technology called “positive reinforcement” comes to the fore here. This term came from psychology and is often used in the context of parenting, where some pleasant consequences for a person or the results of his activities act as positive reinforcement. That is, roughly speaking, the reward for the fact that a person did something right.
It is interesting: Artificial intelligence began to work in a London bar.
The brain, unlike the computer, if possibleto put it, is an "analog device", and biological memory is stable. Analog signals can have several states. While AI built on a binary system can only distinguish between 2 states: “good” and “bad”, our brain is able to interpret what is happening more deeply. The situation can be “very good”, “just good”, “bad” or “very bad”. This principle of work is called “neuromorphic computing” and it is the ability to perform such calculations that will improve the work of AI.
Imagine a baby sitting higharmchair, says Parker. He can swing his arms a lot. In the end, one of these movements leads to some kind of result. Say a child knocks over a cup Suddenly, the neurons that made this movement got an answer and hardened. Thus, a small child learned that the movement of the hand causes an interesting result. This is exactly what neuromorphic calculations are trying to do: teach AI to learn from real experience in the same way as we do.
For this achievement, scientistsdeveloped their own neuromorphic circuits and combined them with nanodevices called magnetic domain wall analog memristors (MAMs). MAMs are very sophisticated devices, but in this case it’s important to know about them: they allow you to create new connections just like in our brains. Thus, it is possible to establish a system of positive reinforcement and begin to train AI.
At the moment, what we have is a littleIt’s like the brain of a real child. Undeveloped and finally not ready to make decisions on their own. Our next step, working with DARPA, is to teach our system to learn something new, not forgetting the previous lessons.
You can discuss this and other news in our chat in Telegrams.