Weird Dreams? That May Be The Point
Updated: May 18
"I dreamed that the lady who ran the organic health food store where my seventh grade creative writing class met to use the computers told me I had to stop coming because she had powerful enemies and when I asked her where I was supposed to buy organic berries now she told me to go to Popeye's and so I walked out of class, heedless of the danger."
Why do our brains create such bizarre stories while we sleep? What in the world is the evolutionary purpose of dreams? Based on the techniques people use to create neural networks, a research assistant professor of neuroscience at Tufts University named Erik Hoel has a new theory: the overfitted brain hypothesis.
Basically, Hoel argues that our brains throw a little strangeness our way so that we can better generalize our daily experiences. “There’s obviously an incredible number of theories of why we dream,” Hoel says. “But I wanted to bring to attention a theory of dreams that takes dreaming itself very seriously–that says the experience of dreams is why you’re dreaming.”
In the world of training AI, a common problem is that the system becomes too familiar with the data it's given and starts to "assume" the training set is the be-all, end-all, containing all the information the AI could ever encounter. This leaves it unprepared for a curveball.
One strategy for overcoming this shortcoming is to throw the data some curveballs of your own; for example, in one regularization method called "dropout," the researchers randomly redact some data. Imagine a self-driving car. Some black boxes appear on its internal screen. The car that encounters this and focuses on the overarching details of its surroundings instead of the specifics of that individual driving experience will develop a better understanding of the actual experience of driving.
Hoel points out that the original inspiration for neural networks was the brain. “If you look at the techniques that people use in regularization of deep learning, it’s often the case that those techniques bear some striking similarities to dreams,” he adds.
Hoel believes that our brains similarly combat our over-familiarity with our own data sets: the information we take in during our day-to-day lives. Under this model, dreams are a way to keep us from feeling like we've totally got the handle on our own realities. Hoel writes, “It is the very strangeness of dreams in their divergence from waking experience that gives them their biological function."
Supporting his theory is research which suggests that the most reliable way to get a person to dream about something that happens in real life to have that individual repeat a new task many many times while awake. Overtraining yourself on a novel task leads to an overfitted brain, he says.
There may be other ways to escape an overfitted brain. Hoel initially created his theory while pondering the neuroscience purpose behind consuming fiction like TV or movies. He holds that these stories can function as, in effect, "artificial dreams," operating as dream substitutions and breaking up the monotony of an all-too-familiar data set. “Life is boring sometimes," he says. "Dreams are there to keep you from becoming too fitted to the model of the world."