The eyes may be the window to the soul, but the pupil is key to understanding how, and when, the brain forms strong, long-lasting memories, Cornell researchers have found.
By studying mice equipped with brain electrodes and tiny eye-tracking cameras, the researchers determined that new memories are being replayed and consolidated when the pupil is contracted during a substage of non-REM sleep. When the pupil is dilated, the process repeats for older memories. The brain's ability to separate these two substages of sleep with a previously unknown micro-structure is what prevents "catastrophic forgetting" in which the consolidation of one memory wipes out another one.
The findings could lead to better memory enhancement techniques for humans and may help computer scientists train artificial neural networks to be more efficient.
The study, published Jan. 1 in Nature, was led by Azahara Oliva, assistant professor in the Department of Neurobiology and Behavior, and Antonio Fernandez-Ruiz, assistant professor and Nancy and Peter Meinig Family Investigator in the Life Sciences, both in the College of Arts and Sciences. The paper's co-lead authors were doctoral student Hongyu Chang and postdoctoral researcher Wenbo Tang.
While the relationship between sleep and memory has been well established, the neural mechanisms that underpin their interplay remain murky. Experiments to elucidate the process generally follow one of two lines: attempting to improve memory retention in sleeping humans, and mechanistic, cellular studies in sleeping rodents. Combining those two fields of research has proved difficult because the subjects are so different. Rodents have been thought to have a very simple, two-step sleep structure - rapid eye movement (REM) sleep, when dreaming occurs, and non-REM slow-wave sleep. However, humans' slow-wave sleep is more complicated, with four stages, the deepest of which is when researchers suspect the brain sifts through and organizes its memories.
Through their joint lab, Oliva and Fernandez-Ruiz used their combined expertise in animal behavior, technology development and computational analysis to see just what goes on in a sleeping mouse's eye - and neurons.
Over the course of a month, a group of mice was taught a variety of tasks, such as collecting water or cookie rewards in a maze. Then the mice were outfitted with brain electrodes and tiny spy cameras that hung in front of their eyes to track their pupil dynamics. One day, the mice learned a new task and when they fell asleep, the electrodes captured their neural activity and the cameras recorded the changes to their pupils.
"Non-REM sleep is when the actual memory consolidation happens, and these moments are very, very short periods of time undetectable by humans, like 100 milliseconds," Oliva said. "How does the brain distribute these screenings of memory that are very fast and very short throughout the overall night? And how does that separate the new knowledge coming in, in a way that it doesn't interfere with old knowledge that we already have in our minds?"
The recordings showed that the temporal structure of sleeping mice is more varied, and more akin to the sleep stages in humans, than previously thought. By interrupting the mice's sleep at different moments and later testing how well they recalled their learned tasks, the researchers were able to parse the processes. When a mouse enters a substage of non-REM sleep, its pupil shrinks, and it's here the recently learned tasks - i.e., the new memories - are being reactivated and consolidated while previous knowledge is not. Conversely, older memories are replayed and integrated when the pupil is dilated.
"It's like new learning, old knowledge, new learning, old knowledge, and that is fluctuating slowly throughout the sleep," Oliva said. "We are proposing that the brain has this intermediate timescale that separates the new learning from the old knowledge."
The fact that pupil tracking during sleep is a noninvasive procedure opens up applications for further study in humans, Oliva said, and could be particularly beneficial for people who have memory deficits associated with mental health conditions.
The benefits even extend beyond humans, into the realm of machine learning, which is woefully cumbersome when compared to the efficiency of the brain.
"The brain can remember a lot of things with a relatively small number of neurons, and how that happens is not understood. How can the brain achieve such a huge feat of memory and cognitive skills with so little resources compared to ChatGPT, which consumes hundreds of thousands of times more energy to do any task?" Fernandez-Ruiz said. "This way of dividing, in time, two key functions of memory is what underlies the huge scope of biological brains for having such amazing memory capacities with relatively low resources. This provides a new opportunity to train artificial neural networks to be more efficient, perhaps, by being more similar to how actual brains work."
Co-authors include Annabella Wulf '24, Thokozile Nyasulu '24 and Madison Wolf '25.
The research was supported by the National Institutes of Health, the Sloan Foundation, the Whitehall Foundation, the Klingenstein-Simons Fellowship Program, and the Klarman Fellowships Program.
