Abstract Perceptual learning alters the representation of sensory input in primary sensory cortex. Alterations in neuronal tuning, correlation structure and population activity across many subcortical and cortical areas have been observed in previous studies. However, relationships between these different neural correlates - and to what extent they are relevant to specific perceptual tasks - are still unclear. In this study, we recorded activity of the layer 2/3 neuronal populations in the whisker primary somatosensory cortex (wS1) using in vivo two-photon calcium imaging as mice were trained to perform a self-initiated, whisker vibration frequency discrimination task. Individual wS1 neurons displayed learning-induced broadening of frequency sensitivity within task-related categories only during task performance, reflecting both learning-and context-dependent enhancement of category selectivity. Learning increased both signal and noise correlations within pairs of neurons that prefer the same stimulus category (‘within-pool’), whereas learning decreased neuronal correlations between neuron pairs that prefer different categories (‘across-pool’). Increased noise correlations in trained animals resulted in less accurate decoding of stimulus categories from population activity but did not affect decoding of the animal’s decision to respond to stimuli. Importantly, within-pool noise correlations were elevated on trials in which animals generated the learned behavioral response. We demonstrate that learning drives formation of task-relevant ‘like-to-like’ layer 2/3 subnetworks in the primary sensory cortex that may facilitate execution of learned behavioral responses. Significance Statement We found that cortical plasticity during perceptual learning alters both neuronal tuning and the structure of pairwise correlations such that they become increasingly aligned to task-related categories, indicating the formation of ‘like-to-like’ subnetworks in layer 2/3 of sensory cortex. Category-specific increases in signal and noise correlations were induced by learning and only observed during active task performance, which points to top-down feedback as a driver of task-related subnetworks.