AI must be given ability to ‘sleep and dream’ just like humans to avoid ‘catastrophic phenomenon’, scientists warn
ARTIFICIAL intelligence can learn better if it dreams and sleeps like a human, according to researchers.
Scientists are hoping to create AI that can replicate this human behavior so it can learn to do tasks better than models that don't take time for sleep.
A team of researchers, including Professor Concetto Spampinato from the University of Catania in Italy, outlined this theory in a .
"We propose Wake-Sleep Consolidated Learning (WSCL), a learning strategy leveraging Complementary Learning System theory and the wake-sleep phases of the human brain to improve the performance of deep neural networks for visual classification tasks in continual learning settings," they wrote.
The team was hoping to avoid a phenomenon called “catastrophic forgetting."
That occurs when an AI model that has been taught new tasks forgets everything it learned.
This can happen if an AI is being trained to learn something additional to a task it's already proficient in.
The researchers conducted an experiment to demonstrate the "importance of dreaming" and noticed "significant performance gain."
Just like humans form long-term memories in their sleep, AI was found to remember tasks better after a rest phase.
The models were trained on sets of data during an “awake” phase before going into periods of “sleeping."
During the sleep period, the AI models were shown images of what they had learned to help them remember.
This was said to be similar to humans during a dreaming phase.
Despite the positive results, some experts don't think mimicking the human brain is the way forward for AI development.
According to , expert Andrew Rogoyski said: "The human brain should not be regarded as the ultimate architecture for intelligence.
"It’s the result of millions of years of evolution and an unimaginably wide range of stimuli.
"We may develop AIs that have structures completely different from their biological designers."