Visual Expertise Depends on How You Slice the Space

Previous studies using fMRI have found that the Fusiform Face Area (FFA) responds selectively to face stimuli. More recently however, studies have shown that FFA activation is not face-specific, but can also occur for other objects if the level of experience with the objects is controlled. Our neurocomputational models of visual expertise suggest that the FFA may perform fine-level discrimination by amplifying small differences in visually homogeneous categories. This is reflected in a large spread of the stimuli in the high-dimensional representational space. This view of the FFA as a general, fine-level discriminator has been disputed on a number of counts. It has been argued that the objects used in human and network expertise studies (e. g. cars, birds, Greebles) are too "face-like" to conclude that the FFA is a general-purpose processor. Further, in our previous models, novice networks had fewer output possibilities than expert networks, leaving open the possibility that learning more discriminations, rather than learning fine-level discriminations, may be responsible for the results. To challenge these criticisms, we trained networks to perform fine-level discrimination on fonts, an obviously non-face category, and showed that these font networks learn a new task faster than networks trained to identify letters. In addition, all networks had the same number of output options, illustrating that visual expertise does not rely on number of discriminations, but rather on how the representational space is partitioned.