Item

Abstract

Ambiguous stimuli constitute a powerful method to dissociate between the physical
properties of the stimuli and their representation in the brain. Following this idea, we applied a new computer-vision algorithm based on Support-Vector-Machines (SVMs) to create three-dimensional morphed faces (linear interpolated) between humans and monkeys in order to investigate how species-dependent face information is encoded in the inferior-temporal (IT) cortex of the macaque brain. Previous psychophysical experiments using these stimuli have shown that human subjects tend to classify ambiguous morphs as discrete instances of the human/monkey categories (‘categorical perception’). Moreover, subjects draw the category boundary closer to their own species (at approximately 60%human/40% monkey).
We recorded the single-unit-activity (SUA) of 118 neurons and the local field potential
(LFP) at 58 sites of the IT cortex of one macaque monkey during fixation of these
morphed stimuli. Out of a total of 118 single units, 85% were visually responsive,
23% were selective to faces, 12% selective to monkeys and 14% to humans,
according to standard criteria. To analyze the population activity, we trained different
classifiers (k-Nearest Neighbor, Support vector Machines, K-Means) to learn the
representation (SUA and LFPs) of human and monkey faces and tested them with
the ambiguous stimuli. We found that, symmetric to the findings in humans,
ambiguous faces are categorized by the pattern classifiers in a manner implying a
categorical representation of the faces. Furthermore, the classifiers drew the
category boundary closer to the monkey category (at approximately 40%human/60%
monkey) for both kinds of neural signals.
In contrast to the linear change of the morphed faces, our preliminary results showed that the neural representation of the species information is nonlinear. This nonlinearity suggests an ‘own-species’ advantage in the encoding of face stimuli.
Consistent with learning theories, this advantage seems to be better reflected in our data by a sharper tuning of the monkey-selective cells compared to the humanselective, and not by a difference in the number of cells.