The Weirdness of Our Dreams Could Explain Their Function, Suggests New Theory
Falling from a great height, naked public speaking and simultaneously losing all your teeth. The content of our dreams is often quite weird. While extensive research has teased out some of the likely reasons why we sleep, why this ubiquitous activity should be peppered with strange and nonsensical imagery has remained mysterious. Now, a new theory that takes inspiration from artificial intelligence has suggested that the weirdness of dreams is essential to their purpose.
The theory appears in an article authored by Tufts University Research Assistant Professor and novelist Erik Hoel, published May 14 in the journal Patterns.
From Freud to AI
In the 120 or so years since Sigmund Freud released his influential tome on dreaming, Die Traumdeutung (The Interpretation of Dreams), neuroscience has replaced his outlandish Oedipal theories with several other proposals for dreaming. But we are still some distance from pinning down a concrete explanation for why dreaming occurs. “The structure of dreams is so unlike waking experience and neuroscientists are very used to thinking about waking experience, which seems very obvious and purposeful, but dreams almost seem purposeless,” Hoel tells Technology Networks. Accounting for this seeming purposelessness has produced several theories of dreaming. These include the idea that dreams help us selectively forget undesirable connections, that they help us prepare for problems we are yet to encounter in the real world and perhaps the currently dominating theory, that dreams help us recall and consolidate memories.
But to Hoel, theories that position dreams as a way of revisiting memories have a structural flaw: dreams almost never resemble what our waking life is like. “Dreams themselves never actually involve explicit memories that we have,” says Hoel. “The only instance it seems where people really repeat specific memories during dreams is when it is a sign of diagnosable post-traumatic stress disorder (PTSD).” Hoel’s theory, instead of viewing the surreal nature of our dreams as a flaw of the unconscious mind, sees it as central to their function.
He draws inspiration from deep neural networks, algorithms used widely in computer science that function a bit like our own brains. These networks are trained on chunks of trial data that they can then build off to generalize rules about the wider world – think cars trialled on Californian roads to develop autonomous driving technology – but can run into a common problem with this approach. Hoel named his theory, the overfitted brain hypothesis, after this common issue.
The overfitted brain
Neural networks, if trained repeatedly on a limited set of trial data, struggle when presented with test data that differs from the trial. Think of an autonomous car trained only on US roads being subsequently asked to navigate the densely packed streets of New Dehli. To avoid overfitting, AI researchers toy with training data, introducing corruptions and occasional blanks in the data, a technique called dropout, that reminds the network that the exact detail of the training data isn’t perfect and can’t be fully relied on in test scenarios.
To Hoel, our own brains become “overfitted” on the “training data” of our everyday routines, and dreams are our mind’s way of shaking things up. “Humans and other animals, particularly mammals, often learn too well. You can never shut off the process in your brain,” says Hoel. “Even when you are watching TV, the brain is learning.” The overfitted brain, says Hoel’s paper, might still be able to learn and memorize things, but struggles to generalize that information. He gives the example of someone learning a new video game. The player might grasp the basics quickly but then plateau, only to find their performance starts to improve again after a night’s sleep. Could this be due to more than a good rest – could the brain be benefiting from weird dreams that give the brain a new perspective on the rote mechanics of a novel task?
Answering that question, says Hoel, will be a challenge. “Our current neuroimaging methods are not nearly as good as they really need to be to adequately distinguish between all the hypotheses. But you could at least do things that support one hypothesis more than another.” Such a task might separate out the “memory consolidation” hypothesis from Hoel’s overfitted brain theory by testing whether dreams improved memorization or generalizable skills.
Summarizing, Hoel explains that if his hypothesis is borne out, researchers might be able to conduct an entirely different practical approach to studying dreams: “It would be possible to do dream substitutions, that is, to artificially create the effect of dreams.”
Hoel proposes a test where a sleep-deprived volunteer is shown weird, dreamlike visuals through a VR simulation for 20 minutes: “If that can recover a little bit of the effect of sleep loss, then we’d know that maybe dreams are doing something similar.”
Hoel E. The overfitted brain: Dreams evolved to assist generalization. PATTER. 2021;2(5). doi:10.1016/j.patter.2021.100244