The Diversity Problem in Face Research

Source: Photo by Vlad Hilitanu on Unsplash.

This post was co-authored by Clifford I. Workman, Ph.D.

Psychology and allied disciplines, like social and cognitive neuroscience, have problems with WEIRD-ness—that is, they rely too heavily on participants from Western, Educated, Industrialized, Rich, and Democratic societies.

Joseph Henrich exposed the pervasiveness of this problem in 2010, arguing that “scientists now face a choice – they can either acknowledge that their findings in many domains cannot generalize beyond [WEIRD subpopulations] (and leave it at that), or they can begin to take the difficult steps to build a broader, richer, and better-grounded understanding of our species.”

The problems with WEIRD-ness can begin in the laboratory, long before researchers seek out the first volunteers for their studies. This problem is especially true for face research, whether focused on describing the mechanisms enabling face perception, underpinning judgments of attractiveness, extracting socially relevant information from facial cues, or even on devising machine learning algorithms to create and detect faces.

Face researchers often require photographs of faces to use as stimuli in their experiments or to train and test machine learning algorithms. Fortunately, plenty of freely available face databases for researchers are available from which to choose. Unfortunately, many of the best-known databases depict people whose demographics skew WEIRD. The FACES database, for instance—which has been cited more than 800 times according to Google Scholar—comprises only white faces (Ebner et al., 2010). For the sake of comparison, the “Multi-Racial Mega-Resolution” (MR2) face database, which is comprised of European as well as African and East Asian faces, has fewer than 80 citations (Strohminger et al., 2016).

An overabundance of face databases featuring mostly (and sometimes exclusively) white faces (e.g., DeBruine & Jones, 2017; Ebner et al., 2010; Lundqvist et al., 1998) means that relying on word of mouth and even on the literature to find such databases may lead researchers to settle for WEIRD face images despite the availability of more diverse alternatives like the MR2. Similar concerns have been raised about the age makeup of available databases, with many including only images of young faces (e.g., Chelnokova et al., 2014; DeBruine & Jones, 2017b, 2017a; Langner et al., 2010). If we are to take the “difficult steps to build a broader, richer, and better-grounded understanding of our species,” as Henrich urged, a crucial first step is for face researchers to choose faces that are diverse in terms of racial/ethnic background and age.

Connecting Researchers to Diverse Face Databases

In a recent paper published in Methods in Psychology, we describe a novel tool for connecting face researchers to the image databases best suited to their research: the “Face Image Meta-Database,” or “fIMDb” for short (Workman & Chatterjee, 2021). The fIMDb is a “meta-database” that indexes known face databases, their characteristics, and how to access them. The fIMDb enables users to build custom searches that, importantly, can be used to filter out stimulus sets with limited diversity. The fIMDb also allows user submissions of new databases and revisions to existing databases. To date, the fIMDb links to 127 different sources for faces that together contain over 4 million images.

Race and ethnicity and age aren’t the only factors that are poorly represented in most face databases. Up to now, there haven’t been any face databases available to research (of which we are aware) that include faces with visible differences like scars, port wine stains, and other craniofacial anomalies. This inclusion is important because people with facial anomalies are, according to our research, subject to an “anomalous-is-bad” bias that is linked to neural activity in a brain region called the amygdala linked to affective learning, prosocial behavior, and psychological tendencies related to empathy and beliefs about justice (e.g., that people get what they deserve; Workman et al., 2021). We built a tool called the ChatLab Facial Anomaly Database (CFAD) to fill this gap (Workman & Chatterjee, 2021). The CFAD contains 4,351 images of 163 people with visible facial anomalies before and, whenever available, after surgical intervention to limit the visual salience of any anomalies (Jamrozik et al., 2019 & Workman et al., 2021).

In publishing the CFAD alongside the fIMDb, we hope that people who research the psychology and neurobiology of face processing might be inspired by the availability of the CFAD to advance our understanding of the social penalties of looking different. We hope that, together, the fIMDb and CFAD help face researchers to make their research a little less WEIRD.