ENCODING MANIFOLDS ORGANIZE NATURAL SCENE RESPONSES IN MOUSE CORTICAL VISUAL AREAS

Encoding manifolds organize neural response properties globally to an ensemble of stimuli. Unlike standard decoding manifolds, in which points are stimuli embedded in neural response space, encoding manifolds treat individual neurons as points, with nearby neurons exhibiting similar temporal responses across the stimulus ensemble. We construct encoding manifolds from the responses of individual neurons to grating stimuli in the Allen Institute Visual Coding–Neuropixels dataset. Manifolds derived from primary visual cortex (VISp) and five higher visual areas (VISam, VISal, VISpm, VISlm, and VISrl) are continuous and exhibit smooth coordinates corresponding to known response properties, including orientation selectivity and firing rate. Surprisingly, despite being constructed solely from grating responses, these encoding manifolds also capture information about neuronal responses to natural scenes. To quantify relative stimulus preference, we project onto the manifold the log ratio of natural scene to grating responses, computed from either mean or maximum firing rates. This ratio is organized along a global manifold coordinate that separates neurons according to their preference for the two stimulus classes. Layer-specific analyses reveal a preference for gratings is concentrated in layer 6, whereas preference for natural scenes tends to be higher in layers 2/3 and 4. Additionally, neurons tuned to low (0.02 cpd) and high (0.32 cpd) spatial frequencies show stronger preferences for natural scenes than those tuned to intermediate spatial frequencies (0.04–0.16 cpd). Together, these results demonstrate that encoding manifolds uncover a structured relationship between responses to artificial and naturalistic stimuli.