Clarke Was Right, Artificial Intelligences DO Dream
In the excellent 1984 film 2010 (taken from the equally excellent 1982 sequel novel 2010 by Arthur C. Clarke), the SAL 9000 computer asks if it will dream if portions of its memory are interrupted or temporarily deactivated.
Meanwhile, back in the real world, artificial intelligence researchers have found that states resembling sleep-like cycles in neural networks can quell instability that arises with uninterrupted self-learning sessions.
“We study spiking neural networks, which are systems that learn much as living brains do,” said Los Alamos National Laboratory computer scientist Yijing Watkins. “We were fascinated by the prospect of training a neuromorphic processor in a manner analogous to how humans and other biological systems learn from their environment during childhood development.”
Watkins and her research team found that the network simulations became unstable after continuous periods of unsupervised learning. When they exposed the networks to states that are analogous to the waves that living brains experience during sleep, stability was restored. “It was as though we were giving the neural networks the equivalent of a good night’s rest,” said Watkins.
Readers may also remember an Isaac Asimov short story from 1975, Point of View, in which a big computer named Multivac is having problems coming up with the right answers. A dad remarks to his child that Multivac is not a simple mechanical device (a machine) that would be easy to fix, nor is it as intelligent as a man, in which case you could ask what was wrong. His son remarks that maybe Multivac was like a kid:
"... you say you've got to keep Multivac busy day and night. A machine can do that. But if you give a kid homework and told him to do it for hours and hours, he'd get pretty tired and feel rotten enough to make mistakes, maybe even on purpose.
So why not let Multivac take an hour or two off every day with no problem-solving - just letting it chuckle and whir by itself any way it wants to."