Summary: In both social circles and neural pathways, sticking with what is familiar feels safe, but it may be creating a “death spiral” for new information. A new study uses a new theoretical framework to show how Hebbian learning, the principle that “neurons that fire together, wire together,” actually prevents activity from spreading.
While positive reinforcement strengthens existing bonds, it also traps ideas in tight loops. Conversely, “negative” reinforcement or weakening connections is what allows information to break free and explore new areas of a network.
Key Facts
- The Hebbian Loop: The study integrated the rule that repeated interactions strengthen links. Surprisingly, they found that the stronger a connection becomes, the more it acts as a barrier to outside information, keeping activity trapped in a “feedback loop.”
- The Ant Mill Effect: Lead researcher István Kovács compares positive feedback to “ant mills,” where ants follow pheromone trails in a circle until they die of exhaustion. In social or neural networks, positive reinforcement can create similar “death spirals” where ideas just circle back to the same people or neurons.
- Efficiency Through Weakness: For an idea, infection, or signal to spread efficiently, the system must avoid old paths. Weakening existing connections (negative reinforcement) forces the activity to find and “explore” new nodes.
- Universal Dynamics: The model applies to any system where activity propagates, including social media echo chambers, the spread of viral infections, and the way signals travel through the human brain.
Source: Northwestern University
Sticking with the same people might feel safe and comfortable.
But a new Northwestern University study suggests it can actually trap new ideas and behaviors inside tight echo chambers. By contrast, the research shows that when interactions shift away from familiar contacts — and toward new ones — activity can spread more widely.
To explore how activities spread across networks, physicists developed a new theoretical framework that includes simple “learning” rules. While traditional network models assume relationships do not change, the new model shows what happens when connections change with experience. As interactions strengthen or weaken relationships, they gradually reshape the entire network.
The findings not only apply to ideas moving through social networks but to a wide range of systems where activity spreads, including infections passing among people, signals traveling through the brain and behaviors proliferating through groups of animals. Ultimately, the study suggests that whether something spreads or stalls may hinge on a simple choice: revisit the same connections or explore new ones.
The study appeared online today (April 27) in Communications Physics, a Nature Portfolio journal.
“Learning and adaptation are intrinsic to biological and social systems, but understanding the effects of learning remains mostly unexplored in even simple models,” said Northwestern’s István Kovács, who led the study. “We wanted to investigate the impact of learning on network dynamics. We found that positive incentives can strengthen existing connections, which, surprisingly, prevents activity from spreading. When connections weaken, however, the system avoids old paths and can lead to more efficient spreading.”
An expert in complex systems, Kovács is an assistant professor of physics and astronomy at Northwestern’s Weinberg College of Arts and Sciences and a member of the Northwestern Institute on Complex Systems and of the NSF-Simons National Institute for Theory and Mathematics in Biology. Will Engedal, a recent graduate from Kovács’ research group, is co-first author of the paper.
‘Fire together, wire together’
In the new study, Kovács and his team set out to explore Hebbian learning, a simple principle that describes how connections strengthen through repeated use. First proposed by psychologist Donald Hebb in 1949, the concept helps explain how the brain learns from experience and forms memories.
“Hebbian learning is often summarized as ‘neurons that fire together wire together,’” Kovács said. “It means that when two neurons activate at the same time, the connection between them strengthens, making it more likely they will activate together again in the future.”
The team incorporated simple Hebbian learning rules into a network model. In traditional models, nodes (representing people, neurons, animals or other objects) connect to each other with links. While activity spreads along those links, the connections do not change. By incorporating learning into the model, connections change based on positive or negative experiences.
Using the new model, Kovács and his team tested two types of learning: positive reinforcement and negative reinforcement. When interacting nodes received positive reinforcement, they were more likely to interact again. Over time, these connections strengthened. When nodes received negative reinforcement, however, they were less likely to interact with each other. These connections weakened over time.
Emergent behaviors shifted depending on whether the source, the target or both nodes learned from the interaction, the researchers found.
Stuck in a ‘death spiral’
When positive reinforcement occurred at the source node, activity circled back along the same routes, becoming trapped in tight loops rather than reaching new areas. But when connections weakened, activity spread outward to explore new paths.
“It’s similar to what happens in the ant mill phenomenon,” Kovács said. “Blind fire ants follow pheromones. But they can accidentally go in a loop. As they follow the loop, the pheromone scent gets stronger, so they continue to follow the same circular trail. The same type of ‘death spiral’ can happen in our model with positive feedback.”
Because the model focuses on a fundamental mechanism — how past interactions shape future ones — Kovács expects the results to hold across many types of spreading processes. Next, his team plans to test whether these learning-driven effects show up in real-world networks and how they interact with more complex, realistic behaviors.
Funding: The study, “Activity propagation with Hebbian learning,” was carried out in collaboration with the HUN-REN Wigner RCP in Hungary and supported by Hungary’s National Research, Development and Innovation Office (award number K146736), the National Science Foundation (award number PHY-2310706), the Hungarian Academy of Sciences and the Baker Program of Undergraduate Research at Northwestern University.
Key Questions Answered:
A: Not necessarily in a social sense, but in a network sense, “agreeability” creates echo chambers. If you only talk to people who agree with you, your ideas never leave that circle. To spread an idea, you need the “friction” of new, less-familiar connections.
A: It explains how habits and “thought loops” form. When neurons fire together repeatedly, they create a very strong, efficient path. While this is great for memory, it makes it harder for the brain to integrate new, contradictory information because the signal prefers the “well-worn” path.
A: Yes. When content gets stuck in a saturated loop of the same people resharing it, it loses its “momentum” to jump to new clusters. The most successful spreading processes are those that can break out of their initial “comfortable” networks.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by our staff.
About this neuroscience research news
Author: Amanda Morris
Source: Northwestern University
Contact: Amanda Morris – Northwestern University
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Activity propagation with Hebbian learning” by Will T. Engedal, Róbert Juhász & István A. Kovács. Communications Physics
DOI:10.1038/s42005-026-02638-z
Abstract
Activity propagation with Hebbian learning
Biological and social systems, including infection spreading, inter-regional brain activity propagation, and population spreading, exhibit learning across a broad range of scales. These applications of the contact process therefore call for an extension that incorporates local learning rules.
Here we introduce learning as a positive (Hebbian) or negative (anti-Hebbian) reinforcement of the activation rate between a pair of sites after each successful activation event.
We show that Hebbian learning leads to a rich class of emergent behaviors, where local incentives can produce opposite global effects. In general, positive reinforcement causes the loss of the active phase, while negative reinforcement can turn the inactive phase into a globally active phase.
Our analytical and numerical results demonstrate that, in two dimensions and above, the effect of negative reinforcement is twofold: it promotes the spreading of activity while simultaneously generating effectively immune regions, leading to the emergence of two distinct critical points.
By contrast, positive reinforcement can give rise to Griffiths effects with non-universal power-law scaling, a manifestation of the ‘ant-mill’ phenomenon.
