Neuroscience data is inherently multiway (neurons × time × trials), yet standard pipelines flatten them and discard separable structure. We adopt a tensor-native workflow for cortical population analysis that preserves this organization using Canonical Polyadic (CP/PARAFAC) decomposition and CP-structured regression. Applied to calcium-imaging recordings from mouse visual cortex, CP factorization yields interpretable neuron, time, and trial components without vectorization. We use this framework to investigate a feature-centric hypothesis of plasticity—namely, that the learning drive acts at the level of neuronal features (cells/components) rather than stimulus labels. It carries no stimulus identity; instead, it selects and strengthens neurons that best support the learned representation while weakening others.