Fixational Eye Movements Reframed as a Source of Visual Signal
For decades, a great deal of cortical variability during visual fixation has been treated as noise to be averaged away. This preprint argues for a different interpretation: in awake marmosets, the tiny movements that continue during fixation appear to account for most of the response variability in foveal primary visual cortex, and they do so in a way that is tied to the retinal input itself rather than to an abstract internal fluctuation in the brain.
The work centers on active sensing, the idea that perception is not built from static snapshots alone. Even when gaze seems stable, the eye is never perfectly still. Those small drifts and tremors continually alter the image projected onto the retina, changing which edges, contours, and spatial frequencies land on photoreceptors from moment to moment. In this study, that ongoing retinal motion did not merely introduce scatter into the neural response. It structured the response in a way that can sharpen spatial coding.
One sentence captures the core claim. Fixational eye movements were not just tolerated by the visual system; they helped generate the visual code.
Marmoset Foveal Cortex and Repeated Natural Images
Investigators measured eye movements with high precision while recording neuronal populations in the foveal primary visual cortex of awake marmosets. The animals viewed repeated flashed natural images, a design that allowed the team to separate trial-averaged responses from response fluctuations linked to eye position and motion. Because the same images were shown repeatedly, any systematic change in neural firing could be tied to the interaction between the stimulus and the ongoing movement of the eyes.
The central result was strong and direct. Fixational eye movements accounted for most conditional firing-rate variability and nearly all shared variability not explained by the trial-averaged response. In practical terms, the population activity that might otherwise be dismissed as unstable was largely predictable once gaze dynamics were taken into account.
Retinal Input Carried More Weight Than Extra Eye-State Signals
The analysis then asked whether cortical neurons required explicit information about eye position or eye velocity beyond the retinal consequences of those movements. An image-computable model captured the modulation largely from gaze-dependent retinal input alone. Adding explicit eye-position and eye-velocity signals contributed little predictive power.
That finding matters because it shifts the explanatory burden back onto the retinal image itself. If the stimulus projected onto the retina already contains the relevant consequences of fixational motion, then the cortex may not need a separate internal bookkeeping signal to explain the observed variability. The movement changes the image, and the image changes the spikes. The neural system is responding to a moving sensory pattern, not simply to a still scene with extraneous motor noise layered on top.
Motion Reweighted Spatial and Temporal Image Content
Under natural viewing, retinal motion redistributed image power across temporal frequencies. As the image slid across the retina, components that were spatially organized began to engage temporal channels, and the effects became larger when the resulting motion interacted with spatiotemporal tuning. This is a biologically plausible mechanism for why apparently tiny movements can have outsized consequences: early visual neurons are built to respond to structure across both space and time, so even micro-motions can transform what is represented.
The study’s modeling suggests that this interaction is not an incidental artifact. The response changes follow the tuning properties of the visual pathway. In a system designed to detect contrast, orientation, and motion-related changes, retinal jitter can enrich the effective sampling of a scene rather than merely corrupt it.
Information Per Spike Increased With Successive Model Stages
Model-derived populations provided the clearest functional payoff. Fixational motion increased the spatial information carried per spike relative to stabilized input, and the gain grew across successive nonlinear stages. The progression is important because it implies that the benefit is not confined to a single processing layer. As visual signals move through nonlinear transformations, the coding advantage can accumulate.
This is not a minor efficiency tweak. It suggests that the visual system may exploit eye movements as a source of deliberate sensory diversity, extracting more usable information from the same scene. In a stabilized condition, the image becomes easier to describe but potentially less informative. With natural fixational motion, the system receives a richer stream of retinal samples that can improve the precision of spatial representation.
Noise, Reframed as Beneficial Reafference
The paper’s conceptual reach lies in how it reclassifies a familiar phenomenon. Activity often assigned to cortical noise is not only structured visual reafference generated by the observer’s own movement; in this account, it can be beneficial signal that sharpens spatial representations. That distinction changes how variability should be interpreted in sensory neuroscience. What looks like instability at the level of single trials may actually be the signature of a sensing strategy optimized for extracting detail from the environment.
In the context of vision, this is a particularly strong argument. The eye is not a camera fixed on a tripod. It is an active sampling organ, constantly probing the world through minute movements that keep the retinal image alive, dynamic, and informative. The data here support the idea that the visual cortex is adapted to exploit that motion, not just endure it.
What the Preprint Establishes and What It Does Not
The study is still a preprint and has not been peer reviewed, so its conclusions remain provisional. The reported findings are nevertheless internally coherent: eye motion explains a large share of the observed variability, retinal motion can account for that effect with limited need for extra eye-state terms, and the modeled consequence is improved spatial information per spike.
The work does not settle every issue around how general this mechanism is across visual areas, behavioral states, or stimulus conditions. It also leaves open how these dynamics compare with other forms of sensory noise, and how they should be incorporated into broader theories of population coding. Even so, the paper makes a forceful case that fixational motion should no longer be treated as a nuisance variable at the margins of visual processing. In this dataset, it is part of the signal itself.