Primates can learn to associate sensory cues with particular movements according to arbitrary rules. We used positron emission tomography (PET) to study the neural network involved in learning such arbitrary associations by trial and error. Ten subjects were scanned at four different stages of learning a visuomotor conditional task (VC). The subjects were required to associate four different visual patterns, presented one at a time, with four different finger movements. Scan 1 was acquired during initial learning. Scans 2, 3 and 4 were performed after further interscan training periods of 1, 3 and 5 min. In order to control for non-specific time effects that could have confounded the learning-related rCBF changes, we also acquired four sensory-matched control scans, in which no movements were performed. In order to evaluate changes over time that were specific to learning the association of visual cues with movements, we acquired four scans during the learning of a motor sequence task. The statistical model tested with SPM considered both main effects of tasks and task x time interactions independently for each of the three experimental conditions. The right lingual gyrus and the left parahippocampal cortex increased their activity over scans in the VC task as compared to the sensory control. The right inferior frontal sulcus, the body of the caudate nucleus and a left cingulate motor area were specifically implicated in learning the VC task, showing task x time interactions with the motor sequence task. These findings suggest that the learning process involves a distributed network in the ventral extrastriate and prefrontal cortex, in association with the basal ganglia and the parahippocampal gyrus.