Perceptual learning improves sensory sensitivity through practice but is typically specific to trained stimuli. Two contrasting accounts have been proposed: fine-tuning of feature-selective neurons in early sensory cortex, which raises a stability-plasticity dilemma, and reweighting of stable sensory inputs by higher-level readout. To reconcile these views, we recorded neuronal populations in macaque visual area V4 during training on a camouflaged collinear contour detection task. Behavioral improvements were strictly orientation specific, yet learning-related neural enhancements were unexpectedly widespread, regardless of neuronal tuning. Critically, these enhancements were confined to late, feedback-influenced response components, whereas early feedforward activity remained stable. These results support a feedforward-gated feedback mechanism: Learning potentiates higher-order representations that deliver diffuse feedback, broadly recruiting early cortical neurons, while behavioral specificity arises because this modulation is gated by feature-selective feedforward inputs associated with the trained stimulus. Rather than being sequestered in a single locus or process, perceptual expertise emerges from hierarchical, bidirectional interactions that enable learning while preserving core sensory stability.