fb-pixel Skip to main content

Why are dreams so weird?

Occasional doses of nonsense can help computers make sense of the world. Maybe something similar is happening with us.

Jonathan Borofsky hangs one of the figures in his "I Dreamed I Could Fly" installation at the Museum of Fine Arts.

Library Tag 10222000 ARTS
Jonathan Borofsky hangs one of the figures in his "I Dreamed I Could Fly" installation at the Museum of Fine Arts. Library Tag 10222000 ARTSGREENE, Bill Globe Staff

Dreams are peculiar. They’re more often salads of mundane daily activities than absurdist romps. That said, they are not just like reality. The phone might have no buttons. The person who’s your dad is not really your dad. And of course, all too rarely, you might be able to fly.

This odd quality leads one neuroscientist to suggest that the purpose of dreams — if there is one — may be to keep our brains on their toes. In an opinion piece in the journal Patterns a few weeks ago, Erik Hoel, a theoretical neuroscientist at Tufts University, described how some computer neural networks seem to need periodic doses of nonsense to stay flexible and perform well. Are brains, he wonders, doing something similar when they dream?


For decades now, research scientists have been exploring whether dreams serve a purpose. (In case you’re wondering, Sigmund Freud’s ideas in “The Interpretation of Dreams,” poetic though they are, have not stood the test of time.) One school of thought is that dreams may be a side effect of other neurological processes of clear importance to survival, namely thinking and sleeping, but they may not have any significant adaptive function of their own. Another school of thought is that dreaming may aid in memory consolidation and learning. People who’ve been given a memory task and dream about it seem to improve their performance the next time they perform the task more than people who don’t dream about the task. It’s not clear, though, if the dreams themselves are helping or if they’re just a byproduct of whatever brain process is at work improving performance.

In his new article, Hoel takes inspiration from neural networks. These are computer programs whose structure loosely resembles that of neurons in the brain and that are capable of impressive feats of pattern recognition. Systems involving neural networks underlie everything from Google Translate to Amazon recommendations. But if some advanced networks, called deep neural networks, train on the same data over and over, they can get stuck on the details and become worse at generalizing, or seeing the big picture.


“It’s like a self-driving car just memorizing the routes it’s supposed to drive,” Hoel says. If it can’t generalize to a wide variety of roads and routes, “it’s not really self-driving, then, is it?”

Researchers address this problem with neural networks by feeding them training data that’s missing big chunks of information or is otherwise contradictory, essentially forcing them to work harder to find meaningful patterns. Perhaps, Hoel suggests, the not-quite-right quality of dreams sharpens the brain’s ability to generalize about situations instead of having to memorize specific responses to them.

Anna Schapiro, a professor of psychology at the University of Pennsylvania, says that in broad strokes, this phenomenon in deep neural networks may resemble what happens with dreams. However, it’s unclear how laboratory experiments could be designed to differentiate this hypothesis from others about dreaming’s purpose.

That said, variation — weirdness, essentially — is widely thought to be important in the brain’s methods for consolidating memories.

“The purpose of memory is not to just record something so you can play it back like it happened,” says Erin Wamsley, a professor of cognitive neuroscience at Furman University, in South Carolina. Whether the learner in question is a neural network or a human toddler, the ability to use memories for learning depends on deducing larger patterns that organize the information. “You want to use all your past experiences to see the commonalities,” she says.


Existing research suggests that when we reactivate our memories, we aren’t repeating the original experiences exactly as they were. “If you dreamed about this phone conversation, it would not resemble the actual phone conversation. It would be really different, and weird, and bizarre,” Wamsley says. “But we know that that actually is how memory so-called replay is. It’s not exact.”

If something as central as memory relies on putting some noise in with the signal, perhaps it’s not so surprising that dreams, whatever their role, remix our daily experiences too.

Veronique Greenwood is a writer whose work has appeared in such publications as The New York Times, The Atlantic, and National Geographic.