Disentangling Stimulus & Population Dynamics in Mouse V1: Orthogonal Subspace Decomposition for Neural Representation

Abstract

Understanding how the primary visual cortex of mice represents the external sensory input separately from the internal states is a fundamental challenge in systems neuroscience. Our work contributes to the problem of decoupling the stimulus-driven and internally generated components of neural activity in the primary visual cortex. Internally generated (or intrinsic) activity refers to neural dynamics that are not directly driven by sensory stimuli, reflecting the brains ongoing, endogenous processes. Neuronal activity encodes both external stimuli and internal cortical states. The internally modulated activity, though not directly observable, can be inferred from the shared structure in population responses, and thus, serves as a proxy for the internal cortical state. We developed a two-phase Partial Least Squares Regression (PLSR) framework that decomposes neural activity into two orthogonal low-dimensional subspaces: (1) a "population" sub-space capturing global variability shared across neurons, and (2) a "stimulus" subspace containing dimensions that discriminate between stimulus conditions while being linearly uncorrelated with the population subspace. We focus on the granular (L4) and supragranular (L2/3) layers of awake mice exposed to visual stimuli consisting of optical flow directions, using mesoscopic two-photon calcium imaging. In both L4 and L2/3 layers, many components individually yield above-chance decoding accuracy, yet a small low-dimensional subspace preserves nearly the full decoding performance of the high-dimensional population. Stimulus-driven components exhibit strong cross-mouse correlations, indicating a conserved coding scheme present in both L4 and L2/3. These components are stable across the entire recording session, reflecting robustness of the underlying representation over time. Removing the global modulation did not abolish stimulus discriminability in either layer, suggesting that information about stimulus direction is not dependent on this global signal. Both L4 and L2/3 stimulus components exhibit comparable decoding performance as well as similar tuning representations, suggesting common encoding of stimulus direction across layers.