Researchers taught artificial intelligence to “sleep”

The international group of scientists forced the impulse neural network to alternate training with the “dream” to perform two different operations without the rewriting of the connections received during the first task. Motherboard writes about this.

“Now there is a huge tendency to use the ideas of neurobiology and biology to improve the work of machine learning algorithms. Sleep is one of them, ”said the co-author of the study and specialist of the University of California at San Diego Maxim Bazhenov.

Artificial neural networks often reach superhuman heights. However, when it comes to consistent training or solving one problem after another, they are not able to gain new knowledge without losing old.

“After high -quality training, it is very difficult to teach to perform a completely new operation. And if this still succeeds, then the old memory will be damaged, ”said the co -author of the study and researcher at the Czech Republic of the Academy of Sciences Pavel Sanda.

He explained that in neuromir this action is called “catastrophic forgetting”, which can only be solved with the help of “consolidation of memory”. This is a process that helps transform recent short -term memories into long -term, often occurring during periods of quick sleep.

According to the scientist, the reorganization of memory plays a big role in why people generally need to sleep. If this process works incorrectly or interrupted, then a person may have The fall of Bitcoin: serious mental disorders.

“This phenomenon can be observed in older people who tell in detail situations from childhood, but with difficulty [remember] that they ate for lunch yesterday,” Sanda noted.

Researchers used previous achievements in the field of plasticity of memory and modeling sleep. They used the neural network to simulate sensory processing and training with reinforcement in the brain of the animal.

Scientists set the model two separate tasks in which she trained to distinguish between punishment and remuneration.

Then they checked whether the AI ​​will demonstrate “catastrophic forgetting”. As a result, each training session on the second task destroyed the knowledge gained in the first.

But, forcing the algorithm to imitate biological sleep through the activation of artificial neurons in the form of noise, scientists discovered progress. The rapid alternation of phases supposedly relaxed with training on the second task allows the model to “remember” how to perform the first.

“This is another good demonstration that straightforward principles can lead to not such simple effects,” Sanda said.

Recall that in September the British startup created a machine learning model to drive different types of cars.

In August 2021, DeepMind has developed a universal architecture of PERCEIVER IO for processing all types of input and output data.

In February, Sonantic created AI-Algorithm, imitating flirty speech patterns thanks to new “non-western sounds”, including sighs, pauses and light laughter.

Subscribe to FORKLOG news at Telegram: Forklog AI – all news from the world of AI!