Sparse Coding of Faces in a Neuronal Model: Interpreting Cell Population Response in Object Recognition.

Abstract

Response to faces as measured by cell discharge in the temporal cortex of monkeys suggests a sparse cell-population coding of complex visual stimuli. The prevailing view assumes that a sparse population code requires the joint contribution of a relatively small group of cells (a neuronal ensemble) for effective coding and recognition. This assumption is based primarily on the consistent observation that single cells in the temporal cortex are broadly tuned rather than narrowly tuned to individual faces. It has been argued that the joint activity of a relatively small number of broadly tuned cells, each responsive to a different constituent feature of a face, could form an ensemble code selective enough to distinguish individual faces. In the present study, schematic faces were presented as stimuli to a model neuronal system for visual pattern learning and recognition. This model effectively codes individual faces by means of competitive activity among single cells during recognition instead of by ensemble coding. The computer simulation permitted an analysis of the activity profiles of all tuned cells during learning and recognition of the faces. All cells were found to be broadly tuned even though coding was mediated by the discrete output of single cells on a competitive basis in a sparse neuronal population rather than by the joint activity of a group of cells. The results show that the observation of broad tuning of cells in temporal cortex under typical experimental conditions does not warrant the conclusion that neuronal ensembles are required for the coding of individual faces. Suggestions are made for changes in the design of experiments to better test hypotheses about the coding of faces (or any other complex visual patterns).