It always seems to us that artificial intelligence algorithms do certain tasks not only faster than humans, but much faster. But paradoxically, the more artificial intelligence resembles the human brain, the better it works.
A team of engineers at IBM have built a physical neural network that literally resembles human neurons. When engineers tested the AI algorithm on a new type of wiring, they found that it performs tasks with the same efficiency as conventional neural networks, but uses 100 times less energy.
If this neural wiring takes root, then soon artificial intelligence will be able to do much more calculations, spending much less energy. The problem is that computer chips and neural network algorithms speak two different languages, and as a result, they are slower. But in the new system, hard and software fit together perfectly, and therefore the new AI system completes all tasks faster without compromising accuracy.
This is a step towards building silicon-based neural networks. Usually, such AI systems did not work as well, but the new study simulated two types of neurons: one was responsible for fast computation, the other was created to store long-term information.
There are, of course, some doubts about the claims of researchers who argue that the answer to creating a working artificial intelligence is to recreate the human brain. By and large, we still don't know how the brain works. And there is a possibility that much of what is in our brain will be useless for the computer.
Nevertheless, new research shows that the physical structure of human neurons can provide a lot of useful things for artificial intelligence algorithms, and in modern conditions, energy conservation is an important condition for the optimal functioning of any system.