Hidden Motion Maps Reveal a Secret Code in the Brain’s Visual Cortex

For decades, neuroscience students have learned a tidy story about the primary visual cortex, the brain region that serves as the first cortical stop for visual information. Each neuron there, the story goes, is a specialist: it fires most vigorously when a visual pattern matches its preferred orientation, and it stays loyal to that preference no matter what else is happening in the scene. This simple population code, in which features of the world are mapped neatly onto the cortical surface, has anchored textbook models of how the brain represents what we see. A new computational study, published in PLOS Computational Biology, shows that this tidy rule breaks down dramatically when it comes to one of the most behaviorally important aspects of vision: motion.

The research team, led by Jonathan David Touboul together with Hannah Germaine, Isaiah Choi, Stephen D. Van Hooser, and Jérôme Ribot, set out to ask how neurons in the primary visual cortex, or V1, represent stimulus direction and temporal frequency, the two key ingredients that together define how fast and in what manner a visual pattern moves across the eye. To answer the question, they assembled an unusually rich dataset of cat V1 responses to visual stimuli in which orientation, direction of motion, spatial frequency, and temporal frequency all varied together. That deliberate co-variation matters. Most earlier experiments changed one stimulus attribute at a time, which made it impossible to see whether a neuron’s preference for one feature actually held steady when other features changed. By sweeping through combinations of attributes, the researchers could test the classical invariance assumption directly, and they found it wanting.

The first surprise concerned temporal frequency, which describes how rapidly a moving pattern flickers or cycles past the eye. When the team mapped preferred temporal frequency across the cortical surface, they found the map was strikingly uniform: unlike the famous pinwheel maps of preferred orientation, which rotate and fracture in beautiful patterns across V1, temporal frequency preference looked essentially the same from one cortical location to another. On the surface, that might seem to suggest that temporal frequency is encoded in a bland, global way, with no spatial structure worth studying. But the second and bigger surprise showed why that conclusion would be premature.

When the researchers examined preferred direction of motion, they discovered that in more than half of the cortical territory they analyzed, a neuron’s preferred direction reversed depending on the temporal frequency of the stimulus. A site that preferred rightward motion at one temporal frequency would flip to preferring leftward motion at another. This is a direct violation of the classical doctrine of invariance, which holds that a neuron’s preferred features should remain stable regardless of the rest of the stimulus. Instead of a single, fixed direction preference, V1 turned out to contain four distinct map motifs, recurring architectural patterns defined by how direction preference does or does not flip as temporal frequency changes. The functional architecture of V1, long portrayed as a static atlas of fixed preferences, is in fact a dynamic landscape whose structure depends on the stimulus itself.

Here is where the study makes its most provocative leap. If direction preferences are unstable and temporal frequency preferences are spatially uniform, how does the brain manage to represent motion at all? The team turned to convolutional neural networks, the same deep-learning architectures that have transformed computer vision, and trained them to decode stimulus properties from patterns of V1 activity. The results were striking: direction, temporal frequency, and even stimulus speed, the ratio that combines spatial and temporal frequency, could be decoded accurately from V1 responses at all cortical locations they examined. Information about motion was not confined to special patches of cortex or to neurons with textbook-tuned properties. It was available everywhere, distributed across the tissue in a form that no single classical tuning curve could explain.

The authors interpret this as evidence for a latent encoding scheme. Rather than storing motion information in conspicuous, spatially organized maps, V1 appears to embed it in subtle, spatially distributed modulations of activity, small shifts in the pattern of responses across many neurons that individually look unremarkable but collectively carry precise information about how a stimulus is moving. This kind of code is invisible to the traditional experimental approach of measuring one neuron’s preferred stimulus and assuming that preference is fixed. It only becomes visible when you analyze population activity as a whole, which is precisely what the convolutional network approach was built to do. In effect, the study suggests that machine-learning decoders can read information out of V1 that the classical population-code framework does not even predict should be there.

The implications ripple outward in several directions. For one, they complicate the long-standing analogy between V1’s orientation map and a sensory atlas in which each cortical location stands for one fixed feature of the world. If direction preference flips with temporal frequency across more than half of V1, then the meaning of activity at a given cortical location depends on stimulus context, not just on where the activity occurs. That context dependence echoes themes from other sensory systems, where cortical responses have increasingly been shown to depend on stimulus history, behavioral state, and natural scene statistics. The new work adds a specific, quantitative demonstration that even the most basic tuning properties of V1 neurons, the ones measured in countless introductory experiments, are not as invariant as the field assumed.

The findings also carry a methodological lesson about how neuroscience should hunt for neural codes. The classical approach, characterizing neurons one at a time with simplified stimuli, was enormously productive, but it was designed to find codes that are local and explicit. A latent code, by contrast, hides in the relationships among responses across space and across stimulus conditions, and it requires analytical tools that can integrate over both. The use of convolutional neural networks as decoders here is part of a broader movement in systems neuroscience: rather than asking what each neuron prefers, researchers increasingly ask what information can be extracted from population activity, and the answers have repeatedly revealed capacities that single-neuron measurements missed. This study applies that logic to one of the most heavily mapped regions in the brain and still finds hidden structure.

There are also practical consequences for how experiments should be designed. Because the team covaried orientation, direction, spatial frequency, and temporal frequency, they could detect interactions that single-parameter designs would have averaged away. The four map motifs they identified only exist in the joint space of direction and temporal frequency; a study that varied direction alone would have reported a stable direction map, and a study that varied temporal frequency alone would have reported a uniform map, and neither would have noticed that the two findings contradict each other under the classical framework. As computational approaches become standard in sensory neuroscience, the expectation is that more supposedly fixed cortical maps will turn out to be conditional on stimulus context, and that more information will be found lurking in distributed response patterns.

What the study does not yet settle is how this latent code is read out by downstream brain areas, or whether animals actually exploit the distributed motion information that decoders can extract. A convolutional network can find decodable structure in V1 activity, but the visual system’s own circuits must do so with biological hardware, and the mechanisms that would make distributed, context-dependent signals useful to later processing stages remain an open question. Still, the core message is clear and likely to resonate widely: the primary visual cortex, the most intensively studied piece of sensory cortex in the mammalian brain, still holds surprises about how it represents the moving world. Motion, it turns out, is not written into V1 as a fixed map of direction preferences. It is written in the fine-grained, shifting pattern of activity across the cortical sheet, a latent code that only modern computational tools have learned to read.

Subject of Research: Latent population coding of stimulus motion direction and temporal frequency in primary visual cortex

Article Title: Latent encoding of movement in primary visual cortex

Article References: Germaine, H., Choi, I., Van Hooser, S. D., Ribot, J., & Touboul, J. D. (2026). Latent encoding of movement in primary visual cortex. PLOS Computational Biology, 22(10), e1014826. https://doi.org/10.1371/journal.pcbi.1014826

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014826

Keywords: primary visual cortex, V1, motion perception, temporal frequency, direction selectivity, population coding, convolutional neural networks, neural decoding, cortical maps, computational neuroscience, cat visual cortex, stimulus speed