All people and creatures with a central neural system need sleep to stay healthy, energized, and simply function properly. We have always assumed that no need for sleep is what distinguishes machines, robots, and AI from us. Recent studies, however, suggest that this assumption might be wrong and sleep can be beneficial for some AI networks.
Researchers from Los Alamos National Laboratory tried to develop neural networks that resemble the way humans learn to see. The network was asked to classify unknown objects. Garrett Kenyon, one of the researchers, compares it to how a kid would approach such a task. For example, if they are given photos of different penguins, rabbits, and antelope and the last animal is unfamiliar to the child, they will still put it in a separate group. That’s what the team wanted to get from AI.
After a long time of learning, the system started glitching and producing random images, something close to a person’s hallucinations. Researchers tried various options to solve the issue and eventually decided to expose the network to noises resembling what our neurons get during deep sleep. This artificial sleep helped to stabilize the system.
It is important to mention that this problem happens to biologically realistic AI networks. According to Garrett Kenyon, it is usual “when training biologically realistic processors, or when trying to understand biology itself.” So, most AI systems are safe and do not need sleep.
Another problem with artificial neural networks is that they can learn only one set of tasks well. Teaching AI additional things is very hard. If it is done, the first set of knowledge is usually damaged because the system overwrites the tasks.
Research published in the journal PLOS Computational Biology argues that artificial sleep can be used to solve the problem. In the beginning, the team trained a spiking neural network (an artificial neural network that resembles the human brain a bit closer) for the first task, then the second, and only after that gave it some sleep. This sequence didn’t work. Researchers had to try a few other orders to discover that the most effective option is to alternate training and sleep sessions while teaching the second task. That is how artificial sleep helps AI networks be lifelong learners like humans and animals.
Okay, we got that sleep can help artificial neural networks function more effectively, but where are such networks even used? For example, in speech recognition. That’s right, these profound conversations you have with Siri or Alexa are possible thanks to artificial neural networks. Recommendations of products that can be interesting for you on Amazon, streaming platforms like Netflix and Hulu with their movie recommendations, and services like Spotify with music you might like all use ANN as well.
Some other examples are recognition of handwritten characters, classification of signatures and assessment of their authenticity, and even detecting cancer cells and analyzing MRI images.
But let’s mention again that the problems described in the article occur in the systems that try to repeat biological processes (like the experiment with animal photos classification), and lifelong learning is not needed in all the networks. So, most systems in the examples wouldn’t need sleep to function properly.
Resources:
https://www.scientificamerican.com/article/lack-of-sleep-could-be-a-problem-for-ais/
https://www.vice.com/en/article/k7byza/could-teaching-an-ai-to-sleep-help-it-remember
https://www.geeksforgeeks.org/artificial-neural-networks-and-its-applications/
https://www.tutorialspoint.com/artificial_intelligence/artificial_intelligence_neural_networks.htm
https://www.xenonstack.com/blog/artificial-neural-network-applications
I could never imagine that such a practice existed. This a very informative article! What a genius idea by Garrett Kenyon to treat a tech that was inspired by a human nervous system as a human itself.
In certain situations, such as when developing physiologically accurate processors or attempting to comprehend biology, I agree that ANNs can profit from simulating sleep. Artificial sleep seems to help ANNs be lifelong learners and better at some tasks, such identifying novel objects or picking up new information without losing what they already know. It’s also intriguing to consider how ANNs are used in a variety of real-world contexts, including voice recognition, product suggestions, and the detection of cancer cells. It is crucial to understand that not all ANNs require sleep simulation in order to work effectively, and that this could only be appropriate in specific circumstances.
First of all, I want to write that this article was really something new and surprising!!! And unexpected. The very fact that artificial intelligence with its development is becoming more and more close to a person is frightening.