There’s nothing better for the human mind than a good bit of sleep, and it turns out that artificial minds can benefit from it as well. In a new study, an AI program was taught to sleep in order to become smarter.
In their current form, AI programs are able to quickly learn single activities. However, they struggle to learn new tasks without forgetting how to do prior tasks. For example, teaching a language model to perform a different activity may make it less effective at understanding text.
A new study believes that allowing an AI program to replicate human sleep will change this. With artificial sleep patterns, programs may be able to retain information better, allowing artificial intelligence to learn more than one task effectively.
Published in science journal PLOS Computational Biology, it was revealed that neuroscience techniques are now being introduced to AI research. Scientists believe that this is the key to creating artificial general intelligence that can be spread across multiple tasks.
Study co-author Pavel Sanda explained: “Once properly trained, it's very difficult to teach them [a] completely new task. And if you succeed in training the new task, you end up damaging the old memory.”
Sanda revealed that AI can suffer from the same mental issues as a real brain, including “catastrophic forgetting” which destroys long-term memory. In the past, it’s been reported that AI can even hallucinate, a side effect of lack of sleep for humans.
The study explains that “consolidation of memory” — turning short term memory into long term — only happens during a human’s REM cycle. With the implementation of REM sleep into an AI’s frequent processes, they may be able to become multipurpose.
Across testing, new training for artificial intelligence resulted in catastrophic forgetting of old activities. However, by incorporating sleep, this was heavily circumvented.
“It's another nice demonstration that very simple principles can produce not so simple effects. We used inspiration from real sleep, but the model is orders of magnitude simpler,” Sanda expanded in a Vice article.
Whether or not this form of AI training becomes commonplace or not is uncertain. Cutting-edge AI like OpenAI’s GPT-4 is still blowing minds without such techniques, but there’s always room to improve.
It just goes to show that there’s a lot of space for artificial intelligence to evolve. How advanced will this technology be ten years from now?