The primate ventral visual system is critical for object recognition, but how its neural representations change during adult categorization learning remains debated. Some studies suggest that neural responses of visual areas change with training, e.g., during perceptual learning. Others propose that behavioral changes are readily explained by adjustments in the readout of information by downstream frontoparietal regions, with sensory regions remaining largely unchanged. Here, we investigated changes in neural representations in macaque visual areas V4 and IT during the learning of novel categorization tasks. Visual stimuli varied along two independent attributes, with a category boundary defined by a combination of these attributes. We used chronic recordings to track neural population responses over multiple days of training as the monkey learned the task through correct/incorrect feedback. Additionally, we recorded from the same populations during a fixation task with the same stimuli. To link behavioral improvement to neural responses, we analyzed changes in single-electrode and population-level activity during learning. As the monkey learned the categorization task, population-level representations in both V4 and IT for the two classes became more separable. This improved separability resulted from changes in single neuron tunings. Interactions between neurons did not enhance separability; rather, they decreased it relative to the expectation from independent responses. Similar changes were observed in both V4 and IT neural populations, but changes were more pronounced in IT and the two areas had notably different response time courses. Overall, our findings suggest that feed-forward responses of the ventral stream are modified during new categorization learning, but involve complex interactions that can reduce linear separability. This interplay helps reconcile previously conflicting results about the role of sensory areas in learning.