Summary: For decades, neuroscientists believed that learning made the brain more efficient by encouraging neurons to act independently—each doing its part with minimal overlap. A new study from the University of Rochester challenges that idea: as people learn a visual task, neurons in the visual cortex become more coordinated, sharing information to form a more flexible, inference-driven representation of the world.
By monitoring the same small networks of neurons in visual cortex over several weeks, the team found that mastery of a visual discrimination task was accompanied by increased shared activity among neurons. Rather than reducing redundancy, learning increased the amount of task-related information that neurons carried in common. This coordinated activity appears to let the brain combine incoming sensory signals with prior expectations, producing perception that is both more robust and more adaptable.
Key Facts
- Coordination increases with learning: As subjects learned to discriminate complex visual patterns, neurons shifted from largely independent responses to highly coordinated activity.
- More shared information: Learning raised the redundancy of information across neurons, challenging the traditional efficiency model that predicted less overlap.
- Engagement-dependent effect: The coordination emerged only when subjects were actively performing the task and making decisions; it vanished during passive viewing of the same stimuli.
- Driven by feedback: The authors link this change to feedback from higher-level brain areas that embed prior knowledge into sensory processing.
- Implications for AI: The results suggest that artificial systems might benefit from generative feedback loops that let internal models shape sensory representations, improving learning from limited data.
Source: University of Rochester
When a skill improves—recognizing a familiar face in a crowd, spotting a typo instantly, or anticipating the next play in a game—sensory neurons in your brain tend to coordinate more closely, sharing information instead of acting strictly independently.
This conclusion comes from research led by Shizhao Liu in the labs of Ralf Haefner and Adam Snyder at the University of Rochester’s Del Monte Institute for Neuroscience, reported in Science. The results challenge a long-standing view in neuroscience that learning enhances efficiency by minimizing redundant signals across neurons.
Liu and colleagues recorded the same populations of neurons in visual area V4 over weeks while subjects learned to distinguish orientations in distinct tasks. Rather than observing decreased shared activity, the researchers saw a clear increase in redundancy: neurons began to share substantial amounts of task-relevant information as training progressed. The coordinated activity was strongest in neurons most relevant to the task and was particularly pronounced during the moments when the subject made a decision.
“The dominant idea has been that learning helps the brain by making neurons act more independently so information can be read out cleanly,” Liu said. “Our data point to a different process: sensory areas actively combine incoming signals with learned expectations, performing inference instead of simply encoding inputs.”
How learning reshapes neural teamwork
Traditional models of sensory processing emphasize a largely feedforward flow of information: early sensory areas pass signals up to higher regions, where behaviorally relevant features are extracted. Under that view, learning reduces correlated variability so each neuron provides unique, nonredundant information.
The new findings align with a generative inference framework, which treats perception as a bidirectional process. In this perspective, sensory neurons represent evolving beliefs about the causes of sensory input; those beliefs are updated by exchanging information between feedforward sensory evidence and feedback expectations. Learning, then, increases the exchange of task-relevant information across neurons rather than eliminating it.
Tracking neurons as learning unfolds
The team used chronic recordings with implanted arrays to follow identical neurons across days of training. At the start, neural responses were largely independent, with minimal redundancy. As training continued, redundancy grew: roughly half of each neuron’s information could be explained by shared activity with other recorded neurons. This increase also evolved dynamically within individual trials over hundreds of milliseconds, consistent with gradual accumulation and sharing of evidence during decision-making.
Crucially, the redundancy increase depended on active task engagement. When the same images were presented without a requirement to respond—passive viewing—the coordinated pattern was absent. This suggests the redistribution of information relies on top-down signals that are engaged when the brain must use sensory data to guide behavior.
Rather than representing a permanent rewiring, these changes appear flexible and context-dependent, guided by recurrent and feedback interactions that allow sensory circuits to switch modes according to task demands.
Insights for health and artificial intelligence
Understanding how learning boosts coordination among sensory neurons may shed light on conditions where perception or learning is impaired. If effective learning depends on neurons coordinating their responses, disruptions to that coordination could underlie some learning disorders—making sensory signals less useful for forming stable internal models.
For artificial intelligence, the findings point toward architectures that incorporate generative feedback loops rather than relying solely on feedforward discriminative mappings. Such two-way systems, where internal models influence how inputs are represented, may learn more quickly from limited data, handle uncertainty more robustly, and adapt flexibly across tasks.
“Most AI today maps inputs directly to outputs,” Haefner said. “Integrating internal models that shape sensory representations could produce systems that generalize better and learn faster from fewer examples.”
Key Questions Answered:
A: The classical efficiency idea treats independence as the best way to maximize unique information. This study suggests a different form of efficiency: by sharing task-relevant signals, neurons create robust, flexible representations that are easier for downstream areas to interpret, especially under uncertainty.
A: If learning depends on neurons coordinating to form shared representations, then disruptions in the feedback or recurrent signaling that fosters coordination could impair the brain’s ability to integrate sensory input with expectations, undermining perceptual learning.
A: It points to “two-way” AI architectures where internal models influence perception. Such designs could make artificial systems faster learners from small datasets and more robust when inputs are noisy or ambiguous.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by staff editors.
About this neuroscience and learning research news
Author: Lindsey Valich
Source: University of Rochester
Contact: Lindsey Valich – University of Rochester
Image: Image credited to Neuroscience News
Original Research: Closed access. “Task learning increases information redundancy of neural responses in macaque visual cortex” by Shizhao Liu, Anton Pletenev, Ralf M. Haefner, and Adam C. Snyder. Science. DOI: 10.1126/science.adw7707
Abstract
Task learning increases information redundancy of neural responses in macaque visual cortex
INTRODUCTION
How does the brain turn sensory input into perception and behavior? The classic view frames perception primarily as a feedforward transformation: early sensory representations are passed forward to higher visual areas where behaviorally relevant information is extracted. Feedback is often considered a secondary refinement that modulates feedforward signals during attention or learning.
An alternative is the generative inference framework, which treats sensory processing as fundamentally bidirectional. In this view, neurons encode beliefs about the causes of sensory inputs that are continually updated by combining bottom-up evidence with top-down expectations.
RATIONALE
Generative inference makes a clear prediction opposite to the classical model: learning should increase the sharing of task-related information among sensory neurons—visible as higher redundancy. The classic model predicts the opposite, that learning reduces redundancy and correlated variability to enhance coding efficiency.
To test these opposing predictions, the authors measured changes in information redundancy in visual area V4 of macaques trained to discriminate orientations across tasks. Neural activity was recorded chronically using microelectrode arrays across weeks of training. Redundancy was quantified by comparing population information with and without correlations present.
RESULTS
At the outset of training, redundancy was minimal, reflecting largely independent neural responses. With learning, redundancy rose substantially: by later stages of training, about half of each neuron’s information was shared with other recorded neurons. Redundancy also grew dynamically within trials over hundreds of milliseconds, consistent with the accumulation and exchange of evidence during decision-making.
Importantly, increased redundancy did not reduce the total information available in the neural population. Instead, individual neurons carried more information while sharing a larger portion of it. The effect was stronger during active task performance than during passive viewing, indicating that engagement and top-down influences are crucial.
CONCLUSION
Learning a perceptual task increased information redundancy among sensory neurons, challenging the conventional notion that learning and attention function primarily to reduce correlated variability. The results support a view of cortical sensory processing as a dynamic inference process in which prior expectations and sensory evidence are integrated through feedback and recurrent interactions. This bidirectional framework better explains how perception becomes more reliable and adaptive with learning.