When animals move through complex visual environments, the brain cannot afford to analyze every detail one by one. Instead, it rapidly extracts the overall structure of the scene—for example, the mean (average) direction of motion across many moving elements. This ability, known as ensemble perception, allows the brain to capture the “gist” of a scene at a glance. Yet where, and how, this statistical summary is computed in the brain has remained unclear.

A research team led by LEE Doyun and KIM Yee-Joon at the Center for Memory and Glioscience within the Institute for Basic Science (IBS) has now shown that this process begins much earlier in the visual system than previously thought.

Co-corresponding author LEE Doyun said, “What is especially striking is that this transformation begins already in primary visual cortex. The brain starts compressing complex sensory input into useful statistical summaries at a very early stage.”

In the brain, visual information is processed step by step along a hierarchy of regions. The primary visual cortex (V1) is the first cortical stage that receives visual input from the eyes and is traditionally thought to process simple features such as edges or motion direction. Further downstream, the posterior parietal cortex (PPC) integrates this information into more abstract representations that are linked to perception and decision-making.

The researchers found that V1 already encodes not only the mean direction of complex motion patterns, but also their variance—how dispersed or uncertain the motion is. This information is then carried forward to PPC, where it is reorganized into more abstract category representations that can guide behavior.

To investigate how the brain extracts these visual summaries, the team trained head-fixed mice to classify random-dot motion stimuli according to their overall direction. Unlike conventional motion displays, in which many dots move coherently in a single direction, the stimuli in this study were designed so that each dot moved in a different direction sampled from a controlled distribution. This allowed the researchers to independently manipulate the mean motion direction and its variability.

The mice successfully learned to group eight possible mean motion directions into two motion categories. Even when the motion of individual dots varied widely, the animals could still categorize the overall direction, indicating that they were not simply following a few prominent local signals. Instead, they were extracting a true statistical summary of the scene.

“We showed that the brain does not process complex visual input by tracking each element individually,” said LEE Young-Beom, first author of the study. “Instead, it extracts stable summary information such as mean and variance to rapidly capture the overall structure of the environment.”

Using miniscope calcium imaging, the researchers recorded neural activity in both V1 and PPC while the mice performed the task or passively viewed the stimuli. At the level of individual neurons, only a relatively small subset showed clear selectivity for the global mean motion direction. At the population level, however, neural activity in both regions robustly encoded the mean motion direction—even though most single neurons did not appear strongly tuned on their own.

The study revealed a clear division of labor across the cortical hierarchy. In V1, population activity encoded both the mean and the variance of motion direction, indicating that early visual cortex already computes summary statistics rather than merely relaying local signals. In PPC, by contrast, the representation shifted toward more abstract, task-relevant category information, suggesting that sensory summaries are progressively transformed into task-relevant signals.

The researchers also found that task demands could reshape early visual representations. During active categorization, the neural representation of mean motion direction in V1 became systematically biased toward the center of the learned category. This suggests that even early visual cortex is not purely stimulus-driven, but can be influenced by learning and behavioral context.

Another notable finding was that seemingly “untuned” neurons still contributed substantially to the population code. Even neurons that did not meet conventional selectivity criteria helped support accurate representation of global motion direction when analyzed collectively, highlighting the importance of distributed population coding in the brain.

Co-corresponding author KIM Yee-Joon added, “Our findings suggest that visual information is progressively reorganized—from summary statistics in early visual cortex to more abstract category representations in higher cortical areas. This provides an important clue to how the brain efficiently makes sense of complex scenes.”

By revealing how the brain converts noisy sensory input into stable statistical summaries and then into abstract category signals, the study provides new insight into a fundamental principle of perception. The findings may help explain how the brain rapidly extracts meaningful structure from complex environments and could also inform future work in artificial intelligence and computer vision.

The study was published online in Advanced Science on March 23, 2026.