US, WASHINGTON (ORDO NEWS) — An artificial neural network that simulates the functioning of the brain suddenly showed a need for sleep. A condition similar to a person’s slow sleep restored her performance.
Recall that they in the most general terms imitate the functioning of neural networks of the human brain. In particular, they also consist of peculiar neurons connected by contacts (synapses). Network training occurs due to a change in the conductivity of these synapses.
Often this is where the resemblance to the brain ends. Usually, artificial intelligence specialists do not set themselves the goal of accurately reproducing the functioning of the nervous system. After all, an airplane does not have to copy the flight of a bird, although it also uses the lifting force of the wing. And ANN does not have to be really like a brain in order to successfully solve the tasks assigned to it.
However, some scientists are interested in the principles of the brain and the ability to simulate them using ANN.
“We are studying the pulse neural networks This system, students as well as the living brain.” – said co-author of the new study Watkins Ching (Yijing Watkins) National Laboratory of Los Alamos, US.
Watkins and her co-authors created a neural network that not only learns to see, but does it in much the same way as a human infant or baby animal.
(Let us explain that full vision is built thanks to practice. An experiment is known in which kittens never learned to see horizontal lines because they did not see them in the first five months after birth).
Watkins and her colleagues found that after a long period of training, the neural network began to work unstable. Scientists have tried many ways to rectify the situation. But an effective solution was found only when weak Gaussian noise was applied to the inputs of the neural network for a long time . After that, the system regained working capacity.
Researchers suggest that it is this signal that the visual neurons receive during slow sleep.
“It was like we [in this way] gave neural networks the equivalent of a quiet night’s rest,” says Watkins.
The authors of the new work emphasize that the need for “sleep” arose in the system due to the fact that its training is very similar to brain activity.
The vast majority of ANNs use other learning algorithms that have only distant similarities with the functioning of the nervous system, and do not encounter such problems.
Now, scientists intend to test their findings on the neuromorphic Loihi chip (the one that recently learned to recognize odors and played the role of a “brick” for the neurocomputer that surpasses the mouse brain). Researchers will connect an artificial eye to it and will train them to process visual information. They hope that by letting the chip “sleep” from time to time, they can improve its performance.
The results of the study will be presented at the conference Women in Computer Vision , which will be held June 14 in Seattle, US.
Contact us: [email protected]