ABSTRACT This project will develop a neural network system to dynamically code facial expression, automating a process which, though widely used, is now highly time, labor, and cost intensive. Automation will make this metric accessible to many more researchers at less cost, while eliminating residual interscorer discrepancies. Measurement of facial expression is valuable in many domains, and has led to unique insights in problems as diverse as human development, sociology, developmental and social psychology. From a practical standpoint, different aspects of expression eluicidate whether a listener is empathetic or hostile (important in politics and business), distinguish abusive from nonabusive caretakers (social work), predict divorce in dysfunctional married couples, and may incriminate dissembling witnesses (forensics). Reliability of facial expression as a gauge of nervous system function has been affirmed through autonomic nervous system indices as well as electromyography and electroencephalography. Initially, automating measurement of facial movement will involve static images of faces of a single subject, which will be preprocessed to normalize size and brightness, compressed with a back-propagation image compressor, and fed to a second neural network whose output will correspond to the "action units" of the widely used "Facial Action Coding System." Recent techniques such as weight-sharing will be used where necessary. Recurrent networks will be used to allow processing of dynamical information, with extension of implementation to multiple subjects. This methodological advance has the potential for a significant contribution to the infrastructure of science in many domains.
