Tufts researcher Erik Hoel draws on machine learning and evolutionary science in his hypothesis about our nightly visits to dreamland
What’s the point of dreams? We all have them, for hours each day of our lives, even if we don’t remember. Plenty of mammals and birds dream, too. When you see Fido sound asleep with his eyes rapidly moving, his paws twitching, you know for sure that he’s in dreamland.
Lots of theories have been offered: dreams are used to regulate emotion, like dealing with fears; to consolidate memory, replaying things from your day to help remember them; to solve, or on the other hand to forget, real-world problems. Another theory suggests they help the brain predict its own future states.
None of those theories seem quite right to Erik Hoel, a research assistant professor at Tufts’ Allen Discovery Center who studies consciousness, modeling the relationships between experiences and brain states.
In his research, Hoel works with artificial neural networks—machine learning. Think of Deep Mind, the Google artificial intelligence program that beat the best human players at the almost infinitely complex Japanese strategy game Go.
It turns out that when such machine learning programs do the same task again and again, they can become “overfit”—able to do that one thing really well, but not to learn lessons and create general knowledge that can be applied to different tasks. To prevent that, programmers often introduce random variables, or noise in the data.
In essence, that’s what Hoel thinks our brains are doing when we dream: breaking the cycle of repetitive daily tasks—filling out spreadsheets, delivering mail, tightening pipe fittings—with an infusion of discord, keeping our brains fit.
Have you ever had a problem that just seemed to defy solution? You think and think, but you remain stuck. Then you go to bed, wake up the next morning, and presto, the solution appears. It might well be, Hoel would say, that your thinking was overfitted for the task—just like a machine learning program in need of disruption.
“This fits with anecdotal reports of plateauing in terms of performance on a task, like a video game, only to sleep and have increased performance the next day,” Hoel says. “There is also the long-standing traditional association between dreams and creativity.”
He recently published a paper on what he calls the overfitted brain hypothesis, and it’s been garnering attention in the press—it was the cover story for a recent issue of New Scientist magazine.
The Evolution of a Purpose
How Hoel came to the theory surprisingly begins not with neuroscience, but fiction.
When he was young, he loved reading. His mother ran the bookstore Jabberwocky in Newburyport, Mass., and he spent a lot of time there, like the proverbial kid in a candy store, immersed in fictional worlds. He always wanted to be a writer, but ended up studying cognitive science at Hampshire College, and went on to get a Ph.D. in neuroscience. (He did become a fiction writer, too: his novel The Revelations will be published by the Overlook Press in early April.)
His focus as a student was consciousness, but his love of reading also made him wonder why people are so drawn to reading novels, “which always struck me from a scientific perspective as kind of a very strange activity,” he says. “Fictions are essentially lies—there’s no such thing as Hogwarts. Harry Potter never went there. It’s the opposite of facts.”
Fiction has all sorts of purposes—aesthetic, emotional, even political—but probably also has an evolutionary role, Hoel says. “I think that one could argue that there is a sort of deep biological need for fictions in humans,” he says. Just look at all the TV shows, novels, movies, and video games we consume for an ungodly amount of our waking hours. Those diversions “actually serve deep down some sort of fundamental purpose,” he says.
He soon started seeing links between fiction and dreaming. Take the short stories by Jorge Borges, Hoel says. They are rife with narrative and yet quite otherworldly at the same time—just like dreams. It made him think there must be some evolved purpose of dreaming, a function seen across many species of animals.
He soon started to research sleep and dreaming. But looking closely at the scientific literature about dreaming, he came away with more questions than answers. One prominent recent theory says dreams are created for memory consolidation; but why, Hoel asks, do the dreams so rarely actually mimic those memories? Another says that dreams are for emotion processing, but there’s little empirical evidence for it.
One fact that many dream theories also overlook is that while reptiles and many other animals don’t dream, mammals and birds apparently do. “Dreaming is so ubiquitous across mammals and even birds that there must be a good reason for it,” he says.
It’s widely noted in neuroscience that many traits are highly conserved, meaning that brains seem to operate in much the same way across the animal kingdom, Hoel says. The human brain, while “basically getting more bang for your buck in terms of space and having some more frontal and prefrontal regions, is not significantly different in its neuroanatomy from a canine’s,” he points out.
It’s also true, Erik Hoel says, that “evolution is a great multitasker, so I’d be surprised if there’s just one absolute reason for dreaming and no other reason.”
So what does differentiate mammals from reptiles? “Mothers,” says Hoel. “When an iguana is born, nature’s basically just booting up the iguana program—almost everything is just innate for them.” Reptiles therefore don’t actively learn. On the other hand, young mammals learn from their moms (and dads, too) as they develop and are cared for.
It’s unsurprising, Hoel says, that the creatures that have to learn to survive have the most pronounced dreaming and signs of dreaming. It’s likely a sign of dreaming’s evolutionary importance for learning—and survival.
The Human Secret
While metaphors of brains as computers is a bit overdone, Hoel says, in this case, reversing the metaphor to say that brains are like neural networks is close to the mark. After all, he says, those neural networks were designed by engineers to mimic human circuitry.
The overlap between how humans dream, and how machine learning experts avoid pure memorization and help programs transfer knowledge from one problem to others “lends credence to the idea that the evolved function of dreaming is for precisely these purposes,” he says. “It seems that the most effective way to trigger dreams about something is to have subjects perform repetitively on a novel task like Tetris, likely because the visual system has become overfitted to the task.”
Sleep is widely known to have a restorative effect—just try going without it for a day or two and see how well you function. Precisely how that works is not completely known. Current thinking is that sleep evolved as some sort of metabolic housekeeping activity—at one stage of sleep, the cerebral spinal fluid essentially flushes waste products through the lymphatic system.
But dreaming seems to happen during other parts of sleep, and apparently occurs more than we realize; we tend to remember our dreams only if we wake in their midst. Hoel’s theory is that dreaming is an exaptation, a trait that evolved for one purpose but later takes on others.
In this case, he says, “sleep evolved for molecular housekeeping purposes, and only when brains had to significantly learn during the organism’s lifetime did the goal of avoiding overfitting and increasing generalization become adaptive.”
Another key feature of Hoel’s theory is that it “takes the phenomenology of dreams seriously.” Our nightly hallucinogenic narratives, “containing fabulist and unusual events,” are exactly what dreams would be if they were fulfilling the role Hoel proposes—adding noise to the thinking system.
“The point of dreams is the dreams themselves, since they provide departures away from the statistically-biased input of an animal’s daily life, which can therefore increase performance,” he says. “It may seem paradoxical, but a dream of flying may actually help you keep your balance running.”
Have you ever had a problem that just seemed to defy solution? You think and think, but you remain stuck. Then you go to bed, wake up the next morning, and presto, the solution appears. It might well be, Hoel would say, that your thinking was overfitted for the task.
And what about dreams that seem to be speaking to us—helping us understand our lives, remember loved ones, or even scare us?
Meaning in dreams, he says, is basically a side effect. “I don’t think dogs are imbuing their dreams with meaning, but they still dream,” he says. “Humans can imbue their dreams with meaning, but dreams should still have a purpose for all mammals who regularly do it,” Hoel says.
It’s also true, he says, that “evolution is a great multitasker, so I’d be surprised if there’s just one absolute reason for dreaming and no other reason.”
Hoel comes back to where he started: fiction. “It is worth considering whether fictions, like novels or films, act as artificial dreams, accomplishing at least some of the same function,” he says.
His theory, he says, “suggests fictions, and perhaps the arts in general, may actually have an underlying cognitive utility in the form of improving generalization and preventing overfitting.”
The tradition of fiction goes back much further than the first novel, he says—maybe to the first storytelling shamans. “Maybe that’s part of the human secret—we export some of our learning finessing outside of the body, so that you don’t have to just do it through dreams,” he says. “You can do it through these artificial dreams that maybe even are more impactful because they’re so well structured.”