Proceedings of the Annual Meeting of the Cognitive Science Society

Predictive processing, or predictive coding,1 is transforming our knowledge of perception (Knill & Richards, 1996; Rao & Ballard, 1999), the brain (Friston, 2018; Hohwy, 2013; Knill & Pouget, 2004), and embodied cognition (Allen & Friston, 2018; Clark, 2016; Gallagher & Allen, 2018; Seth, 2015). Predictive processing is a hierarchical implementation of empirical Bayes, wherein the cognitive system creates generative models of the world and tests its hypotheses against incoming data. It is hierarchical insofar as the predictions at one level are tested against incoming signals from the lower level. The resulting prediction error, the difference between the expectation and the incoming data, is used to recalibrate the model in a process of prediction error minimization. Predictions may be mediated by pyramidal cells across the neocortex (Bastos et al., 2012; Hawkins & Ahmad, 2016; Shipp et al., 2013). Andy Clark has characterized predictive processing as creating a “bootstrap heaven” (2016, p. 19), enabling the brain to develop complex models of the world from limited data. This enables us to extract patterns from ambiguous signals and establish hypotheses about how the world works. The training signals that we get from the world are, however, biased in all the same unsightly ways that our societies are biased: by race, gender, socioeconomic status, nationality, and sexual orientation. The problem is more than a mere sampling bias. Our societies are replete with prejudice biases that shape the ways we think, act, and perceive. Indeed, a similar problem arises in machine learning applications when they are inadvertently trained on socially biased data (Avery, 2019; N. T. Lee, 2018). The basic principle in operation here is “garbage in, garbage out”: a predictive system that is trained on socially biased data will be systematically biased in those same ways. Unfortunately, we are unwittingly trained on this prejudiced data from our earliest years. As predictive systems, we bootstrap upwards into more complex cognitive processes while being fed prejudiced data, spiraling us into a “bootstrap hell.” This has repercussions for everything from higher-order cognitive processes down to basic perceptual processes. Perceptual racial biases include perceiving greater diversity and nuance in the faces of racial ingroup faces (the cross-race effect; Malpass & Kravitz, 1969), misperceiving actions of racial outgroup members as hostile (Pietraszewski et al., 2014), and empathetically perceiving emotions in racial ingroup (but not outgroup) faces (Xu et al., 2009), among other phenomena. They are particularly worrying due to their recalcitrance to conscious control or implicit bias training. We may be able to veto a prejudiced thought (but see Kelly & Roedder, 2008), but we cannot simply modify our perceptual experience at will. Recalcitrant predictions such as this are “hyperpriors” and are unamenable to rapid, conscious adjustment. I begin with an overview of predictive processing. I explain that the same principles that allow us to bootstrap our way into full cognition also allow for biases to develop. These biases include perceptual racial biases, which are visual and affective rather than cognitive. I explain how sampling biases in infancy and emotion perception contribute to perceptual racial biases (although many other factors certainly play a role). Finally, I hypothesize that traditional implicit bias training may not be enough to disentangle the web of hypotheses that contribute to perceptual racial bias.