Disenchantment with Emotion Recognition Technologies: Implications and Future Directions

ABSTRACT: With the development of IoT and Machine Learning, it is now possible to automatically evaluate emotions from various data streams among which facial expressions are one of the most prominent. An exponential number of tech companies are providing commercial systems to infer emotions from facial expressions (Software, API or SDK, see Dupré et al., 2018). As shown in the largest benchmark to date (Dupré et al., 2019), results from these technologies are significantly less accurate than human observers. Criticising not only the algorithms’ performance but also the theory underlying these systems, well known scientists in psychology and computer science have called for a halt to the use of these technologies for significant decisions (e.g., in human resources management). While automatic classifiers of identity from faces have been proven to be gender and ethnically biased, emotion recognition appears to be biased as well. However, some encouraging improvements are suggesting solutions to the frailties in emotion recognition technologies.