Researchers are inspired by human sleep to train an AI capable of remembering processes

November 14 (Portaltic/EP) –

A group of scientists has carried out research in which they study how to train an artificial intelligence (AI) to simulate the functioning of a human brain, and specifically the role of sleep, and to be able to remember activities without reaching catastrophic oblivion.

The authors of this new study, published in the scientific journal PLOS Computational Biologyhave been based on the REM phase, which retains memories and reproduces them during sleep, to apply them in simulated neural networks.

REM sleep is one of the five distinctive phases that the brain goes through when the person sleeps and represents 25 percent of the sleep cycle. Specifically, it happens around 70 minutes after falling asleep.

This phase is characterized by the fact that dreams occur in it and it participates in the memory storage process. In addition, it stimulates the regions of the brain used for learning and helps balance mood.

The researchers recall that artificial neural networks find it difficult to sequential learning, since they overwrite tasks learned during training.

“Once they have been properly trained, it is very difficult to teach them a completely new task and if you manage to train the new task, you end up damaging the old one in memory,” said study co-author and researcher at the Institute of Computer Science of the Czech Academy of Sciences, Pavel Sanda, in statements collected by Motherboard.

In this way, when they have assumed a task and go on to the next ones, they end up forgetting the first one, something that does not happen in the human brain thanks to the REM phase of sleep, which recalls said memories and avoids what is called ‘catastrophic forgetting’. ‘.

This, also known as catastrophic interference, is the tendency to forget totally and unexpectedly. information or processes learned previously sequentially.

The researchers believe that this can be solved with so-called ‘memory consolidation’, a procedure that allows transform recent memories short-term into long-term memories.

Sanda has given as an example of catastrophic forgetting how older people can have detailed childhood memories but find it difficult to remember more recent things, such as what they did or what they ate the day before.

Because sleep enhances learning by allowing the “spontaneous reactivation of previously learned memory patterns,” the researchers believe this technique can be applied to AI. To do this, they are inspired by neuroscience and the functioning of the human brain.

To teach this AI to adopt the behavior of the human brain, this group of professionals has carried out tests with simulated neural networks (SNN). With them, they have simulated sensory processing and have learned how the new learning works using the brain of an animal as a reference.

First, the researchers gave this simulated neural network two tasks and it learned to discriminate between punishment and reward, so they concluded that the network can make decisions autonomously.

Next, the way in which this network reaches the so-called ‘catastrophic oblivion’ has been exposed. To avoid this and resemble its behavior to that of a human brain, they have carried out another experiment.

Thus, they have seen how the neural network is capable of maintaining this knowledge if sleep phases are interspersed between short periods of the task following the initial one (which would be the second), which allows the AI remember how to do the task first.