Is artificial intelligence dreaming of the electric sheep? All living things need sleep: dogs do, birds and dolphins do the same. There may even be lice. Without sleep, people become forgetful, hallucinate, and experience physical and psychological problems. Now, Scientists believe AIs may need sleep too from time to time. Even a dream.

In reality, most AIs can only master a set of well-defined tasks: they cannot gain extra knowledge without losing what they have learned before. The problem arises when you want to develop systems capable of performing so-called “lifelong learning”. This is the way we humans accumulate knowledge to adapt to and solve future problems.

While neural networks are more tireless and sensitive than humans, they become forgetful when it comes to sequential learning or learning one new thing after another. Once trained, it is very difficult to teach them completely new tasks. And if you manage to train the new task, you end up corrupting the old memory.

This activity is called “destructive forgetting”. And experts say it’s a problem that can only be fixed with something called “memory defragmentation”; it’s a process in humans that helps turn recent memories into long-term memories, often during REM sleep.

In this way, artificial intelligence can learn and remember to multitask by imitating what you do. sleep helps us consolidate what we learn during our waking hours In fact, there is a huge trend towards examining some ideas from neuroscience and biology to improve machine learning, and sleep is one of them.

Researchers at the University of California San Diego wanted to further study the phenomenon and trained a network of interconnected artificial neurons, similar to the structure of the human brain, to learn two different tasks without overwriting the connections learned from the first task.

The team first tried to train the neural network on the first task, then added a sleep period to the second task and finally to the end. But they found that this directory still deleted the links learned in the first mission. Instead, experiments showed that it is important to do alternate training sessions and sleep while the AI ​​is learning the second task. This helped solidify the connections in the first mission that would otherwise be forgotten.

What does artificial intelligence dream of?

Inside Other research published in Scientific AmericanResearchers at Los Alamos National Laboratory have discovered that artificial intelligence may need to sleep in order to function properly. Stability was restored when they exposed the networks to conditions similar to the waves that living brains experience during sleep. It’s as if neural networks are taking a nap.

Kenyon and his team made their discovery while working on training neural networks. see objects in a way that resembles people. The networks were instructed to classify objects without any samples to compare with, and they spontaneously began producing hallucinatory-like images. After the nets were allowed the electronic equivalent of the dream, the hallucinations ceased.

Physicist Stephen L. Thaler, head of artificial intelligence company Imagination Engines, warns against taking the term “sleep” literally when applied to AI. When we talk about sleep, we’re talking about a cycle between “chaos” and “calmness.”

AI not only needs to sleep, it can also dream. It is possible for an artificial intelligence to reach new answers by dreaming.John Suit, director of consulting technology at robotics company KODA, explained in this Lifewire article. “This is how people work. We are presented with problems or challenges, we overcome them and learn. If we don’t learn best, we face new challenges that are very similar until we arrive at the best answer. It’s a dream.” This may be the key to achieving this in AIs.”

And what if they get high?

Even sleep may not be necessary for artificial intelligence to change your consciousness. According to a recent article published in the journal Neuroscience of ConsciousnessMedicines may also work. In the study, researchers discussed how psychedelic drugs such as DMT, LSD, and psilocybin can alter the function of serotonin receptors in the nervous system. To see what happens when investigating this phenomenon, they tried giving neural network algorithms virtual versions of the drugs.

Conclusion? AI could be wrong, it seems. The photorealistic results of the networks became distorted blurs, similar to how people describe their journeys under the influence of drugs. In other words, the process of producing images with neural networks visually disturbing in a similar way and its biological counterpart, as well as a tool to illustrate psychedelic experiences.

Yes, the field of artificial intelligence is advancing rapidly and has not yet proven practical for widespread use as it is difficult to train. However, the next big step forward with these methods could be a breakthrough in artificial neural networks, which, as science has found, are more energy efficient than other neural networks.

Images: Pexels / Unsplash