With the materialization and development of artificial neural networks, complex challenges similar to those of human intelligence have emerged, especially in the field of learning and permanent memorization, which is why scientists are finding that, as in humans, the solution may lie in deep rest and sleep.
How can sleep improve the function of neural networks?
By: Gabriel Levy Bravo
A neural network is a type of computer and electronic technology that uses artificial intelligence (AI) algorithms by processing autonomous data, mimicking the biological mechanisms used by the human brain[1].
This technology is part of the so-called machine learning universe, known in scientific jargon as deep learning.
So-called neural networks use nodes that mimic interconnected neuronal synapses in a layered structure similar to the human brain.
The algorithms used in neural networks create an adaptive system that computers use to learn from their mistakes and continuously improve; that is, they avoid repeating a wrong action, learning to solve complicated problems, such as creating a piece of music, classifying data, writing a text, summarizing a document, recognizing faces or processing DNA strands.
Breaking the limits of artificial learning
As evidenced by an article published in NewScientist, most artificial intelligence technologies based on neural networks can only master a limited and narrowly defined set of tasks[2]:
“Neural networks cannot acquire additional knowledge later on without losing everything they had previously learned. The problem arises if you want to develop systems that are capable of achieving so-called lifelong learning.
Lifelong learning is the way humans accumulate knowledge to adapt and solve future challenges.” Pavel Sanda, Czech Academy of Sciences[3].
Albert Sanchis, in a publication in Xataca’s Magnet Blog[4], states that, although artificial neural networks are tireless and in some cases could be more accurate than human ones, when it comes to sequential learning or learning one new thing after another, they become forgetful, i.e. they are surpassed by the capacity we people have to accumulate knowledge. In other words, a neural network, once it is trained to perform a task, it is very difficult for it to learn completely new tasks.
Sanchis states that, if you finally manage to train the new task, you end up damaging the old memory.
The phenomenon explained by Sanchis is known in the computer science community as “catastrophic forgetting” and is a problem that, so far, can only be solved with something called “memory consolidation”[5], a technique that, imitating the behavior of the human brain, helps transform recent memories into long-term memories, a phenomenon that generally occurs during REM sleep, i.e. when we sleep.
Deep sleep for artificial intelligence
Several scientific experiments have shown that artificial intelligence based on neural networks can learn and remember how to perform multiple tasks, if they mimic the way in which sleep helps humans to consolidate what was learned during waking hours, as reported by some experts in an article in Scientific American[6], which explains why many of the principles of neuroscience are being applied in the field of computer science.
“We found that our artificial intelligence neural networks became unstable after continuous periods of learning.
However, when we exposed the networks to states that are analogous to the waves experienced by living brains during sleep, stability was restored. It was as if we were giving the neural networks the equivalent of a good, long nap.”
Garret Kenyon in Scientific American publication [7].
Garrett’s studies conducted at Los Alamos National Laboratory not only addressed sleep as an innovative solution in the learning processes of neural networks, but also documented how the lack of these periods of rest trigger erratic, unstable and confused actions in machines, in the same way that occurs with humans who come to experience hallucinations in the absence of sleep.
“The decision to expose our biologically realistic networks to an artificial analog of sleep was almost a last-ditch effort to stabilize them. The AI was spontaneously generating images that were similar to human hallucinations.
It is for this reason that we experimented with various types of numerical noise, roughly comparable to the static you might encounter between stations while tuning a radio. The best results were obtained when we used noise with a wide range of frequencies and amplitudes.
The noise mimics the current received by neurons in the human brain during slow-wave sleep, which is deep sleep, without which we cannot live.”
Garret Kenyon in Scientific American publication [8].
Physicist Stephen L. Thaler, president of the artificial intelligence company Imagination Engines, cautions against taking the term “sleep” literally when applied to artificial intelligence; rather, it is a simile of the electrical field produced by human brains during the sleep period.
When we talk about sleeping we refer to a cycle between “chaos” and “calm”.
The AI not only needs to sleep, i.e., it not only needs to reproduce this electric field, but it can also dream or produce new spontaneous nodes and it is even possible for an AI to come up with new responses or connections during this sleep in the same way as it happens with the human brain.
“This is how humans work. We are presented with problems or challenges, we overcome them and we learn. If we don’t learn the best way, we face new, very similar challenges until we arrive at the best answer. A dream state may be the key to achieving this in AIs.”
John Suit, Chief Technology Officer, advisor to robotics company KODA, in an article published in Lifewire[9].
In conclusion, in the same way that artificial intelligence based on neural networks has proven to be efficient at learning when it mimics the human synapse, sleep based on an analog reproduction of the human electrical current while we sleep, is becoming the best tool to ensure that this machine learning is deep and long-lasting, accumulating properly with other previous learning.
[1] IBM. (August 17, 2020). Neural Networks. Available at https://www.ibm.com/cloud/learn/neural-networks
[2] New Scientist. (November 10, 2022). AI uses artificial sleep to learn new task without forgetting the last. Disponible en https://www.newscientist.com/article/2346597-ai-uses-artificial-sleep-to-learn-new-task-without-forgetting-the-last/
[3] Op. Cit. Disponible en https://www.newscientist.com/article/2346597-ai-uses-artificial-sleep-to-learn-new-task-without-forgetting-the-last/
[4] Magnet. (November 15, 2022). To be more efficient, artificial intelligences need something suspiciously human: sleep. Available at https://magnet.xataka.com/un-mundo-fascinante/para-ser-eficientes-inteligencias-artificiales-necesitan-algo-sospechosamente-humano-dormir?utm_source=xataka&utm_medium=network&utm_campaign=repost
[5] Vice. (November 11, 2022). Scientist taught an AI to ‘Sleep’ So That It Doesn’t Forget What It Learned, Like a Person. Available at https://www.vice.com/en/article/k7byza/could-teaching-an-ai-to-sleep-help-it-remember
[6] Scientific American. (December 5, 2020). Lack of Sleep Could Be a Problem for AIs. Available at https://www.scientificamerican.com/article/lack-of-sleep-could-be-a-problem-for-ais/
[7] Op. Cit. Available at https://www.scientificamerican.com/article/lack-of-sleep-could-be-a-problem-for-ais/
[8] Op. Cit. Available at https://www.scientificamerican.com/article/lack-of-sleep-could-be-a-problem-for-ais/
[9] Lifewire. (January 18, 2021) Why Artificial Intelligence Needs to Sleep. Available at https://www.lifewire.com/why-artificial-intelligence-needs-to-sleep-5095871