Artificial neural networks play a crucial role in the utilisation and analysis of large amounts of data. One particular field of application for Psychology and Medicine research is the recognition and classification of human facial expressions. The systems used for this are very powerful and are developed as automated processes (end-to-end) for processing large data sets and trained accordingly. However, these black box models often exhibit irregular behaviour when used, especially in the case of subpopulation shifts (mixture shock), which impairs the performance and validity of the predictions.

ComputerExternal link scientists from the Computer Vision Group JenaExternal link, led by Professor Joachim Denzler, have taken up this problem and analysed in a recent study which factors could have an influence on the irregular behaviour of AI models for the recognition of facial expressions. Two common applications (HSEmotion-7 and ResidualMaskNet), which are able to recognise a total of six basic emotions as well as a neutral expression, were selected for this purpose. Together with the Otorhinolaryngology Department under the direction of Professor Orlando Guntinas-Lichius at Jena University Hospital (JUH), 36 healthy volunteers were initially asked to express the various emotions in a random order in front of a camera in a standardised procedure. In order to avoid distortions caused by human attributions, the focus was exclusively on the participants' ability to express facial expressions, which was recorded using frontal shots. In order to take into account the application context in neurological diagnostics, the recordings were made four times with and twice without the applied electrodes of the high-resolution surface electromyography. A second group of subjects consisted of 36 patients with unilateral facial paralysis, the appearance of which can be accepted as having a demonstrable influence on the quality of AI model predictions. These images with the different facial expressions were taken using 3D scans.