'The unbearable (technical) unreliability of automated facial emotion recognition' by Federico Cabitza, Andrea Campagner and Martina Mattioli in (2022) 9(2) Big Data and Society comments
Emotion recognition, and in particular facial emotion recognition (FER), is among the most controversial applications of machine learning, not least because of its ethical implications for human subjects. In this article, we address the controversial conjecture that machines can read emotions from our facial expressions by asking whether this task can be performed reliably. This means, rather than considering the potential harms or scientific soundness of facial emotion recognition systems, focusing on the reliability of the ground truths used to develop emotion recognition systems, assessing how well different human observers agree on the emotions they detect in subjects’ faces. Additionally, we discuss the extent to which sharing context can help observers agree on the emotions they perceive on subjects’ faces. Briefly, we demonstrate that when large and heterogeneous samples of observers are involved, the task of emotion detection from static images crumbles into inconsistency. We thus reveal that any endeavour to understand human behaviour from large sets of labelled patterns is over-ambitious, even if it were technically feasible. We conclude that we cannot speak of actual accuracy for facial emotion recognition systems for any practical purposes. ...
Emotional artificial intelligence (AI) (McStay, 2020) is an expression that encompasses all computational systems that leverage ‘affective computing and AI techniques to sense, learn about and interact with human emotional life’. Within the emotional AI domain (but even more broadly, within the entire field of AI based on machine learning (ML) techniques), acial emotion recognition (FER), which denotes applications that attempt to infer the emotions experienced by a person from their facial expression (Paiva-Silva et al., 2016; McStay, 2020; Barrett et al., 2019), is one of the most controversial (Ghotbi et al., 2021) and debated (Stark and Hoey, 2021) applications.
In fact, ‘turning the human face into another object for measurement and categorization by automated processes controlled by powerful companies and governments touches the right to human dignity’ and ‘the ability to extract […physiological and psychological characteristics such as ethnic origin, emotion and wellbeing…] from an image and the fact that a photograph can be taken from some distance without the knowledge of the data subject demonstrates the level of data protection issues which can arise from such technologies’. On the other hand, opinions diverge among the specialist literature. Some authors highlight the accurate performance of FER applications and their potential benefits in a variety of fields; for instance, customer satisfaction (Bouzakraoui et al., 2019), car driver safety (Zepf et al., 2020), or the diagnosis of behavioural disorders (Paiva-Silva et al., 2016; Jiang et al., 2019). Others have raised concerns regarding the potentially harmful uses in sectors such as human resource (HR) selection (Mantello et al., 2021; Bucher, 2022), airport safety controls (Jay, 2017), and mass surveillance settings (Mozur, 2020). In addition, the scientific basis of FER applications has been called into question, either by equating their assumptions with pseudo-scientific theories, such as phrenology or physiognomy (Stark and Hutson, Forthcoming), or by questioning the validity of the reference psychological theories (Barrett et al., 2019), which assume the universality of emotion expressions through facial expressions (Elfenbein and Ambady, 2002). Lastly, others have noted that the use of proxy data (such as still and posed images) to infer emotions should be supported by other contextual information (McStay and Urquhart, 2019), especially if the output of the FER systems is used to make sensitive decisions, so as to avoid misinterpretation of the broader context. According to Stark and Hoey (2021) ‘normative judgements can emerge from conceptual assumptions, themselves grounded in a particular interpretation of empirical data or the choice of what data is serving as a proxy for emotive expression’. From a technical point of view, FER is a measurement procedure (Mari, 2003) in which the emotions conveyed in facial expressions are probabilistically gauged to detect the dominant one or a collection of prevalent emotions. As a result, FER can be related to the concepts of validity and reliability. A recognition system is valid if it recognizes what it is designed to recognize (i.e. basic emotions); it is reliable if the outcome of its recognition is consistent when applied to the same objects (i.e. a subject’s expression). However, when FER is achieved by means of a classification system based on ML techniques, its reliability cannot (and should not) be separated from the reliability of its ground truth, i.e. training and test datasets (Cabitza et al., 2019). In this scenario, reliability is defined as the extent to which the categorical data from which the system is expected to develop its statistical model are generated from ‘precise measurements’, i.e. human ‘recognitions’ exhibiting an acceptable agreement. This is because, by definition, no classification model can outperform the quality of the human reference (Cabitza et al., 2020b).
In this study, we will not contribute to the vast (and heated) debate still currently going on about the validity of automatic FER systems (Franzoni et al., 2019; Feldman Barrett, 2021; Stark and Hoey, 2021), that is, we do not address the classification task from the conceptual point of view (how to define emotions, if possible at all) nor merely from the technical point of view (how to recognize emotions, whatever they are). For the sake of argument, we assume that the main psychological emotion models make perfect sense and we do not address how robust recognition algorithms are, how well they perform in external settings, and, most importantly, how useful they can be, i.e. whether they provide the benefits that their promoters envision and advocate.
Instead, we focus on the reliability of their ground truth, which is not a secondary concern from a pragmatic standpoint (Cabitza et al., 2020a, 2020b). To that end, we conducted a survey of the major FER datasets concentrating on their reported reliability as well as a small user study by which we address three related research questions: Do existing FER ground truths have an adequate level of reliability? Are human observers in agreement regarding the emotions they sense in static facial expressions? Do they agree more when the context information is shared before interpreting the expressions?
The first question is addressed in the ‘Related work and motivations’ section and the answer is in Table 3. The other questions are addressed by means of a user study described in the ‘User study: Methods’ section and whose results are reported in the ‘Results’ section. Finally, in the ‘Discussion’ section, we discuss these findings and their immediate implications, while in the ‘Conclusion’ section we interpret them within the bigger picture of FER reliability and relate them to implications for the use of automated FER systems in sensitive domains and critical human decision making.
