Item

Abstract

Conceptual models of face recognition have assumed that faces are encoded as points of an abstract face space relative to an average face, or face prototype (e.g. [1]). So far it has been
largely unclear how such a prototype-referenced encoding of faces could be implemented
with real neurons. Recent electrophysiological evidence seems to support the relevance of
prototype-referenced encoding. Neurons in macque inferotemporal cortex, which have been
trained with human faces, tend to show a monotonic tuning with the caricature level of the
stimuli [2]. We present a neural model that accounts for these new electrophysiological results.
The hierarchical model consists of multiple layers of neural detectors modeling properties
of neurons in the dorsal visual processing stream. The rst layer models simple cells using Gabor
lters with with physiologically realistic parameters. A second layer combines responses
of Gabor lters that carry signicant information about a training stimuli into more complex
features. The complex features in the model are based on the Principal Components of the Gabor
responses, which could be extracted using simple Hebbian-like learning rules. The highest
hierarchy layer models neurons in area IT. The responses of these neural detectors increase
monotonically with the distance of the input feature vector, from the previous layer, and the
average feature vector over all training faces. In addition, neural detectors on the highest hierarchy
level show a broad tuning with resepect to the direction of the difference vector between
input feature vector and this average vector.
The model was tested with gray-level images that were generated using a morphable 3D
face model [3]. The model was trained with 98 randomly chosen faces from a data basis with
200 faces. It was tested with caricatures and anti-caricatures of 4 selected faces. In addition we
tested lateral caricatures of the faces, which lie on curves in face space that connect the four
selected example faces. Exactly the same stimuli had been used in the electrophysiological
experiments [2].
After training, a signicant number of the neural units on the highest level of the model
show a monotonic tuning with the caricature level of the faces, and a moderate tuning with
respect to facial identity, consistent with the electriophysiological results. The model provides
a physiologically plausible concrete neural implementation of face spaces. Future work will
explore its computational properties and coding efciency in comparison with classical neural
models for face recognition.