The brain does not wait for sensory information to reach the frontal cortex before making a decision. New recordings show it starts far earlier, and no existing model predicted this

For decades, neuroscience assumed the brain makes decisions at the top. New research found it starts at the bottom.

The standard model of how the brain makes decisions has been taught in neuroscience courses for decades, programmed into artificial intelligence systems for almost as long, and treated as settled enough to build entire fields of research on top of. It goes like this: sensory information enters the brain at the bottom, travels upward through increasingly complex processing layers, and eventually arrives at the frontal cortex, where the actual decision is made. The early sensory regions at the base of this hierarchy do their job and pass the information along. Decision-making happens at the top.

A new study from the University of Illinois Urbana-Champaign has recorded what is actually happening in the brain while decisions are being made, and the picture looks considerably different. The decision-making process does not wait until information has climbed the full hierarchy. It begins at the very first sensory processing layer, earlier than the model predicts is possible, driven by signals flowing back down from higher brain regions before the sensory processing is even complete.

Why the traditional model made sense

The feedforward model of brain computation has intuitive appeal and genuine explanatory power. When you see an object, light hits the retina, the visual system processes edges, then shapes, then objects, then categories, each stage building on the last until the frontal cortex can say: that is a cup. The sequential, hierarchical nature of this process is real and well-documented. And it provided the original inspiration for convolutional neural networks, the deep learning architecture that powers most modern AI vision systems, which process information in exactly this kind of layer-by-layer upward flow.

The problem is that the feedforward model predicts a specific timing: early sensory regions should show sensory activity before higher regions engage, and decision-related activity should appear only after information has traversed the full hierarchy. The frontal cortex decides. The primary sensory cortex records.

Evidence has been accumulating for years that this timing is not always what is observed. Decision-related signals have been found in sensory regions that should, on the feedforward account, be too early in the processing chain to contain them. The explanations offered for these findings have typically been conservative: artifacts of movement, contamination from attention signals, or correlations that are not causally meaningful.

The Illinois study provides a mechanistic explanation for why those signals are there and what they are doing.

Recording the decision as it forms

The research team, led by electrical and computer engineering professor Yurii Vlasov, approached the problem from an unusual direction. Rather than studying higher-order brain regions where decisions were assumed to live, they focused on the very bottom of the sensory hierarchy: the primary somatosensory cortex, known as S1, which receives and initially processes touch and body-position information from the skin and muscles.

To study decision-making in this region, the team placed mice in a virtual reality corridor and trained them to make perceptual decisions based on tactile cues delivered through their whiskers. The whisker system in mice is one of the most precisely mapped sensory circuits in neuroscience, making it an ideal platform for studying how sensory input becomes a behavioral decision. As mice navigated the corridor and made their choices, dense electrophysiological electrodes recorded the firing patterns of neurons throughout the depth of the primary somatosensory cortex.

The neural activity they observed did not match what the feedforward model predicts. During the period of evidence accumulation, when the mouse was processing sensory information and forming a decision, the high-dimensional activity of many neurons simultaneously collapsed into a single variable. This variable then gradually increased across the entire cortical column in a synchronized ramp, tracking the accumulation of evidence toward a decision boundary.

This pattern is the signature of a decision computation, not a sensory relay. And it was happening in the primary somatosensory cortex, the first stop for touch information in the cortical hierarchy, before that information would have had time to travel through the full processing chain and return from the frontal cortex.

The feedback that explains it

The mechanism the researchers identified involves top-down regulation: signals flowing backward from higher brain regions into S1, modulating how that region processes its own sensory input while the decision is still forming. Rather than simply recording touch and passing it along, S1 is being actively shaped by information coming from above, feedback about what the animal expects, what it has decided so far, and what the decision context requires.

This bidirectional communication means that S1 is not a passive sensor but an active participant in the decision computation. The frontal cortex does not wait for S1 to finish its job and send the answer up. It reaches down into S1 while the sensory processing is still happening and begins shaping it toward a decision, creating a loop in which the bottom of the hierarchy and the top are influencing each other simultaneously rather than in strict sequence.

“The neural code of the brain is still mostly an unknown language,” said Vlasov. “But this systems-level understanding can be viewed as a potential impact on how more efficient artificial neural networks can be built, how the next generation of AI can be thought through. Maybe with these analogies that we learn from real brains, we can improve AI further.”

The finding places the Illinois results in the context of a theoretical framework Vlasov calls natural intelligence: the idea that the brain’s remarkable computational efficiency, its ability to perform extraordinarily complex tasks using far less energy than any current AI system, is not incidental but architectural. The bidirectional feedback loops that the study documents may be precisely the feature that allows biological intelligence to do so much with so little.

What this means for AI

The most commercially and technologically significant implication of the finding is what it says about artificial intelligence architecture. The dominant AI paradigm, deep learning with feedforward convolutional networks, was directly inspired by the feedforward model of brain computation. Layer-by-layer processing, each layer building on the one below, no information flowing backward: this is how most state-of-the-art AI systems are structured today.

If the brain’s actual decision-making architecture involves continuous bidirectional feedback between early and late processing layers, with decisions emerging from the dynamic interaction of those layers rather than from a final judgment at the top, then current AI systems may be missing a fundamental feature of biological intelligence. Not just a refinement but a core structural principle.

“We want to learn from a billion years of evolution,” Vlasov said. “How is biological intelligence organized architecturally? Can we learn from the architectural side of the brain and emulate that to make AI more effective, less power hungry, and more intelligent than it currently is? In the level of decision-making, that’s where current AI is lacking.”

The energy dimension is significant. Current large AI models require enormous amounts of electricity to train and run, a constraint that is becoming increasingly relevant as AI systems are deployed at scale. The brain performs its computations on roughly 20 watts, less than a dim light bulb. If the bidirectional feedback architecture is part of what makes biological computation so efficient, then incorporating it into AI design could reduce the energy requirements of artificial systems dramatically while simultaneously improving their performance on exactly the tasks where AI currently struggles: flexible reasoning, decision-making under uncertainty, and rapid adaptation to novel situations.

What the study does not yet establish

The researchers are careful about the limits of what their data can support. The recordings were made in mice, and the decision-making signals in S1 were recorded during a task involving whisker-based tactile perception. Whether the same bidirectional feedback mechanism operates in human sensory cortex during the kinds of decisions that humans face daily, involving vision, language, social reasoning, and abstract thought, remains to be established.

The feedback from higher brain regions to S1 that the researchers propose as the mechanism was inferred from the pattern of neural activity in S1 rather than directly measured in the higher regions that were providing it. Confirming the mechanism fully will require recordings from both ends of the feedback loop simultaneously, tracking the signals as they move in both directions.

The study was also conducted in a single cortical column of a single sensory region. The brain contains many cortical regions, each with its own architecture and connectivity, and the degree to which the findings generalize across them is unknown.

What it does establish is the existence and functional significance of decision-related signals in the primary somatosensory cortex, a region that the dominant model of brain computation assigns no role in decision-making. The signal is there, it tracks the decision as it forms, and it appears before the decision has had time to travel through the full hierarchy and return. That is enough to make the model significantly less complete than it appeared before the recordings were made.

The next step

Vlasov’s team plans to extend the investigation by examining the temporal dynamics of the feedback signals in finer detail, specifically how fast the top-down modulation arrives in S1 and how that timing relates to the behavioral decision the animal makes. They are also developing new neural recording technologies capable of measuring activity across multiple brain regions simultaneously, which would allow the feedback loop itself to be observed rather than inferred.

“By looking at the fast temporal dynamics of neural activity, maybe we can understand better how these feedback loops are engaged in making decisions,” Vlasov said. “Maybe that’s the approach that potentially uncovers these currently unknown mechanisms, how these feedback loops are organized dynamically and how they form and shape different levels of processing. Maybe that can be implemented in new architectures for AI.”

The model that placed decision-making at the top of the brain’s hierarchy was not wrong. It captured something real about how the brain processes information. What the Illinois data shows is that it was incomplete in a way that matters, because the part it missed is exactly the part that may explain why biological intelligence is so much more efficient, flexible, and capable than anything engineers have built so far.

Source

Alex G. Armstrong, Yurii Vlasov. “Neural correlates of perceptual decision-making in the primary somatosensory cortex.” Proceedings of the National Academy of Sciences, 2026; 123 (18). DOI: 10.1073/pnas.2514107123