November 10, 2026
Maureen Searcy
UChicago researchers use machine learning method to enhance memory model
It can take only a few notes to recognize a song or a whiff of familiar scent to transport you back to childhood. On the flip side, sometimes a word or a memory can be right on the tip of your tongue, but you just can’t grasp it. Memory has a funny way of filling in the blanks from the smallest cues or failing you in the most frustrating way.
In a new study published in PNAS, scientists from the University of Chicago and the Max Planck Institute for Neurobiology of Behavior used a memory model to investigate how memory recall can fail. They found that introducing a way to freeze neurons, a method inspired by neuromodulator chemicals in living brains, allowed memories to be retrieved past the point where they normally become inaccessible.
These findings could have implications for further study of living brains, improving memory models to better reflect biology, and providing insight into neural network technology.
The memory landscape
The way memory works in humans is extraordinarily complicated, so scientists study processes like memory formation and recall through simplified mathematical models. These can be run as computer simulations, where researchers can focus on a narrow selection of factors to see how changing certain elements might affect outcome.
One memory model — called the Hopfield network, which was the focus of the 2024 Nobel Prize in Physics — represents a group of connected neurons. Patterns, or memories, are “stored” in the strength of those connections.
Imagine a net made of interconnected strings strung between knots, representing neurons. Certain strings might be tight while others are loose, and those differences in tautness create valleys in the net. Those dips are memories. The act of thinking can be thought of like a ball rolling around the net.
A cue, like a few song notes or a scent, can put a ball close to a memory valley, where it will fall in. Your partial remembrance puts you in the right vicinity, where you can settle into the complete memory.
Yet, when too many memories overlap, where the strings tighten and loosen in many different directions, those valleys can become less defined, flattening out. A cue might put a ball in the vicinity of a memory, but it doesn’t settle into that spot. It might linger before rolling away. Those areas of near-recalls are called “ghosts.”
| Left: A landscape with few “valleys,” where balls have settled into memories (red). Right: A landscape of many overlapping “valleys,” where some balls have settled into a memory pattern (red) while some are adrift in an area with no defined memories (purple). (Illustration courtesy of Daiki Goto) |
The memory cliff
“A neuroscientist named David Eagleman said that ‘the enemy of memory isn’t time; it’s other memories,’” said Daiki Goto, PhD’26, a recent
The Hopfield network has a hard upper limit for memories. Once it holds too many memories, the system collapses, and no memories can be accessed at all. It’s called a “memory cliff,” or catastrophic forgetting.
It is difficult to say whether biological brains have a comparable cliff, but this type of system failure is a weakness of memory models, limiting what such a model can do.
Some artificial neural networks, however, use gates that open and close to control how information flows and how long it gets held. Goto had been studying this gating mechanism — first introduced to him by the late Peter Littlewood — in machine learning models and wondered if it could be used in a memory model to circumvent the memory cliff.
“Neurons are connected through synapses, which is physical wiring, but there are a multitude of chemicals in the brain like neuropeptides, dopamine, and serotonin that send signals and regulate neural activity. This family of chemicals is called neuromodulators,” said Goto. “And some effects of this chemical signaling on neural activity might resemble what gates do in artificial neural networks.”
“It seemed like these two things are very compatible,” said Hector Manuel Lopez Rios, a postdoctoral scholar in Vaikuntanathan’s group. “What would these neuromodulators do to the network? How can this enhance the computability of these networks? Let’s get them together.”
The Vaikuntanathan group teamed up with Monika Scholz, a neuroscientist at the Max Planck Institute for Neurobiology of Behavior who earned her Ph.D. in biophysics from the University of Chicago. Scholz was studying neuromodulation in the roundworm C. elegans, a model organism used for research due to the simplicity of its biological systems.
Combining Scholz’s biological expertise with the Vaikuntanathan team’s experience with machine learning mechanisms, the researchers introduced gating into the Hopfield model to see what would happen once a critical mass of memories was reached. Using theory and simulations, they found that recall remained possible far beyond the memory cliff in an ungated network.
Gating essentially freezes neurons, pinning the ball near a memory, so it stays near the memory rather than rolling away. A trade-off, however, is that those memory patterns are recalled less precisely, and recall works better with stronger cues, like the first several seconds of a song.
The team plans to study the neural activity of C. elegans to help align their gated model to better reflect how neuromodulators work in living brains.
These findings could expand what memory models can do, providing a way to extend or avoid the point of failure. The study could also potentially shed light on how neuromodulators work in biological brains.
“There’s been a lot of experimental work showing neuromodulators to be very important, but on the theory and modeling side, it’s been really difficult to pinpoint the importance through that quantitative eye,” said Lopez Rios. “Maybe these findings could be a steppingstone to address those questions.”
Citation: “Neuromodulation-inspired gated associative memory networks: Extended memory retrieval and emergent multistability.” Goto, Lopez Rios, Scholz, and Vaikuntanathan, Proceedings of the National Academy of Sciences, Oct. 27, 2026.