If an object embedded within noise is encountered repeatedly,our initiali ivague impression will be replaced by a concrete percept.What are the neural correlates of this perceptual improvement? To address this question,here we used fMRI to track the changes of neural activities of the Fusiform Face Area (FFA) when subjects were learning faces.Specifically,in training sessions subjects learned to match a cue face,which was phase-scrambled into three levels of visibility,to unscrambled faces via feedback.In fMRI sessions,stimuli used as the cue faces in the training sessions were presented in an event-related fashion,jittering to different directions,and the subjects reported either their identities (Attend-to-Face) or their motion directions Attend-to-Motion).The fMRI scans were alternated with the behavioral training sessions until the subjects behavioral performance reached asymptote in all three levels.As expected,the accuracy of identifying scrambled faces improved with learning,while the magnitude of the FFA responses elicited by those stimuli increased monotonically.