Summary: A team of biomedical engineers has uncovered a specialized biological circuit that allows the “thinking” parts of the brain to actively reshape how the “sensing” parts process identical visual inputs. By constructing biologically constrained recurrent neural networks (RNNs) and validating the models against in vivo mouse recordings, the team isolated a precise structural mechanism, inhibition-on-inhibition connectivity, that drives cognitive flexibility.
This shift reveals that early sensory areas are highly dynamic, adaptive workspaces actively managed by top-down contextual context, offering a structural blueprint to build significantly leaner, more energy-efficient artificial intelligence models.
Key Facts
- Complicating the Assembly Line: The project expands on the team’s earlier fMRI findings, which revealed that human primary visual areas alter their baseline scanning behavior depending on what rules a participant uses to sort shapes. The identical physical shape sparks completely different visual cortex signatures based purely on task context.
- Biologically Constrained Modeling: Because macro-level fMRI scans are too coarse to isolate individual cell circuits, the team built a simplified recurrent neural network from scratch. Unlike standard, open-ended AI models, this network was strictly constrained to match biology, featuring distinct, dedicated pools of excitatory (firing) and inhibitory (suppressing) digital neurons organized hierarchically.
- The Disinhibition Secret: When the model was trained to switch flexibly between sorting rules, researchers looked inside and discovered it relied on a specific arrangement: inhibitory neurons that specifically suppress other inhibitory neurons (a process known as disinhibition). This specialized loop functions as a bridge, passing top-down instructions from high-level cognitive modules down to low-level sensory entryways.
- The Structural Failure Test: To confirm this circuit was essential, the engineers systematically weakened these inhibition-on-inhibition connections. The model’s ability to switch tasks immediately collapsed. Conversely, weakening other connection types left the network’s processing capabilities completely intact.
- Living Brain Validation: To test the model’s prediction against biology, researchers recorded active neural firing in the visual cortex of living mice. When they selectively silenced the matching inhibitory anchor cells in the living tissue, the animal’s cortex instantly lost its ability to track task context, validating the computational model.
- Redundancy and AI Efficiency: Dr. Rungratsameetaweemana traces her inspiration to working with human patients lacking a hippocampus. Despite missing this critical memory hub, the patients preserved highly flexible cognitive skills. This proved early sensory regions operate with massive functional redundancy, a principle Columbia hopes to use to design lean AI models that match human adaptability on a tiny fraction of today’s large language model energy costs.
Source: Columbia University
Nuttida Rungratsameetaweemana is challenging a story neuroscience has told for decades. According to the conventional account, our eyes collect raw information and relay it through a series of nerves and waystations that lead deep into the brain, eventually reaching the cortex. There, the thinking begins as information is processed and put to use for higher tasks such as reasoning, judgment, and decision-making.
Her group’s work is complicating that account. Last year, the team published fMRI scans showing unexpected levels of activity in the earliest visual areas of the cortex, the regions that first receive visual signals. Rather than passively relaying what the eyes take in, those early areas seemed to process the same information differently depending on what the research participant was doing. When asked to sort shapes by one set of rules, a participant’s early visual system behaved one way. When asked to apply a different set of rules to the same shape, it behaved differently.
In a new paper published today in PLOS Biology, Rungratsameetaweemana and her team at Columbia Engineering show how the brain might pull this off. They built a simple neural network that follows many of the rules that govern real brains. Like the brain, their model contained one class of neurons that drive other neurons to fire and another class that suppress firing.
The team had the model perform a task similar to what the human participants had done while in an fMRI machine. When the researchers looked inside the model to see how the neural network had solved the problem, they found that it relied on one arrangement of digital neurons. Inhibitory neurons that suppress other inhibitory neurons seem to pass key information from the “thinking” part of the system to the “sensing” component of the system.
To test whether that wiring was essential, they weakened those connections in the model, and its ability to switch between tasks collapsed. Weakening other types of connections left performance largely intact. The pattern held up against the living brain as well. In recordings from the visual cortex of mice, silencing the inhibitory cells that anchor this circuit reduced the cortex’s ability to track the task context, just as the model predicted.
We caught up with Rungratsameetaweemana, Maa-Liao Assistant Professor of Biomedical Engineering, to learn more about the research.
This builds on last year’s fMRI study. Why turn to AI models next?
A brain scan gives you a picture of the whole brain, but it’s coarse — you can’t see what individual cells are doing, or what’s happening at the level of circuits. Seeing which regions light up and patterns of activity gave us a reason to go deeper. To get at the mechanism, we needed something we could take apart and change, so we turned to neural models. We build these from scratch, so it’s possible to see exactly what the network is doing to solve the task.
Why keep the models so simple?
If a model has abilities the brain doesn’t have, then anything we find inside it won’t tell us much about real brains. So we did the opposite and built something that only includes features we know to be true about biology. A lot of that builds on earlier work from Tomas Gallo Aquino and Robert Kim, the paper’s co-first authors, among other studies.
We know there are excitatory and inhibitory neurons, so we built those in. We know the brain is organized in a hierarchy, so in the second part of the paper we gave the network two regions: a sensory module that receives input directly, and a higher-level module downstream. Then we can ask how the network uses those ingredients to do the task.
Why do these inhibitory-on-inhibitory connections matter so much?
They give the system very fine control over how information gets represented. These inhibitory neurons turn out to be really important for keeping everything well-controlled, for making sure the right thing is represented in the right way.
There are four kinds of connections you can have between these cells, and the one that matters for this kind of flexible processing is inhibition acting on inhibition. We know it’s important, because the model fails when we take it away. We don’t yet know why it has to be this particular wiring. That question is an important topic for research teams across the world.
How did you get started with this line of research?
Back in 2015, I began working with patients who are missing the hippocampus, the region that lets us form and hold onto memories. If the brain were truly modular, losing that middle piece should leave you unable to do a great many things. But it doesn’t.
These patients can still do all sorts of tasks. That was the first real evidence, to me, that the early regions of the brain are doing more than just relaying sensory information. And it points to something useful: if information is stored redundantly, you can lose a part and still get by. I think that’s how the brain actually works.
What could it mean for AI?
Compare the brain to something like ChatGPT or a large language model. We can do far more, across far more situations, on a tiny fraction of the energy — and without being trained on the whole internet. The brain got there through evolution, through the redundancy built into its wiring.
Our models are recurrent neural networks, which are quite different from the transformers behind today’s large language models. The goal is to work out these principles one by one and use them to make AI leaner and more adaptive. This inhibition-on-inhibition motif is one of them.
What’s next?
We’ve gone back to humans. We’re working closely with clinical collaborators who monitor epilepsy patients with electrodes placed deep inside the brain, letting us record neural activity directly while those patients perform cognitive tasks. These fine-grained measurements will give us data to test our hypotheses against real neural activity.
Funding: This work was funded by the ARL Human Guided Intelligent Systems grant (W911NF-23-2-0067) and the Strengthening Teamwork for Robust Operations in Novel Groups (STRONG) grant (W911NF-22-2-0148).
Key Questions Answered:
A: The traditional story treated your brain like an assembly line: your eyes capture raw pixels, pass them to the visual cortex at the back of your head to outline shapes, and then hand that sketch up to the “thinking” areas at the front of the brain to make decisions. Dr. Rungratsameetaweemana’s team proved that this assembly line is actually a highly interactive two-way street. The thinking parts of your brain don’t wait for information to arrive; they send top-down instructions back down the line to the very first visual areas, actively changing how those raw pixels are processed depending on what task you are currently trying to accomplish.
A: Think of your brain cells as having two basic controls: gas pedals (excitatory neurons that make things fire) and brakes (inhibitory neurons that slow things down). An inhibition-on-inhibition connection, often called disinhibition, is when one brake pad clamps down on another brake pad. By braking the brakes, you temporarily let the gas pedal floor it in a highly controlled, specific way. The Columbia team discovered that this precise double-brake mechanism is the exact cellular pathway the front of the brain uses to send contextual memos back to the visual cortex, allowing the system to instantly pivot its strategy when rules change.
A: Current state-of-the-art AI systems like ChatGPT are computational giants; they require massive, warehouse-sized data centers and staggering amounts of electrical energy to perform tasks because they are trained on nearly the entire internet. The human brain, by contrast, can navigate thousands of complex, fast-changing daily situations on a tiny fraction of that power—roughly the energy needed to run a dim laptop lightbulb. By mapping out the brain’s internal tricks, like the efficiency of the inhibition-on-inhibition motif and structural redundancy, engineers can build compact, recurrent neural networks that match human adaptability without needing massive data scales.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by our staff.
About this visual neuroscience and cognitive flexibility research news
Author: Mohamed Abdelfattah
Source: Columbia University
Contact: Mohamed Abdelfattah – Columbia University
Image: The image is credited to Neuroscience News
Original Research: The findings will appear in PLOS Biology
