A mechanistic account of bodily resonance and implicit bias

Abstract

Implicit social biases play a critical role in shaping our attitudes towards other people. Such biases are thought to arise, in part, from a comparison between features of one's own self-image and those of another agent, a process known as 'bodily resonance'. Recent data have demonstrated that implicit bias can be remarkably plastic, being modulated by brief immersive virtual reality experiences that place participants in a virtual body with features of an out-group member. Here, we provide a mechanistic account of bodily resonance and implicit bias in terms of a putative self-image network that encodes associations between different features of an agent. When subsequently perceiving another agent, the output of this self-image network is proportional to the overlap between their respective features, providing an index of bodily resonance. By combining the self-image network with a drift diffusion model of decision making, we simulate performance on the implicit association test (IAT) and show that the model captures the ubiquitous implicit bias towards in-group members. We subsequently demonstrate that this implicit bias can be modulated by a simulated illusory body ownership experience, consistent with empirical data; and that the magnitude and plasticity of implicit bias correlates with self-esteem. Hence, we provide a simple mechanistic account of bodily resonance and implicit bias which could contribute to the development of interventions for reducing the negative evaluation of social out-groups.

Keywords: Embodiment; Implicit association test; Implicit bias; Neural network model; Self-image; Virtual reality.

Fig. 1

The Self-image Network Model of Bodily Resonance. (A) In this example, the self-image network is comprised of four neural sub-populations that respond to the perception of male (M), female (F), brown (Br) and blonde (Bl) haired features either in the agent itself or in other agents. (B) When the agent perceives its own features, a neuromodulatory signal allows synaptic connections between active sub-populations to be strengthened by a Hebbian learning rule. In this case, the simulated agent is a brown-haired female, such that the self-image network comes to encode strong associations between neurons encoding brown-haired and female features. (C, D) During subsequent perception of another agent, sensory input to sub-populations encoding the features of that agent generates additional activity in the network via recurrent synaptic connections if those features overlap with the encoded self-image. The total firing rate output of the self-image network, likely equivalent to the BOLD response observed in fMRI, can therefore be interpreted as a measure of bodily resonance. In this example, perception of a brown-haired agent produces activity in the sub-population encoding female features, while perception of a blonde-haired agent produces no additional activity in the self-image network. (E) Hence, the degree of overlap between the features perceived in another agent and the encoded self-image correlates with both the total output firing rate of the self-image network and the firing rate of any single unit encoding a feature that is part of the self-image, but not currently being perceived.

Fig. 2

The Drift Diffusion Model of Behavioural Performance in the Implicit Association Test. (A) Each IAT trial consists of a fixation period followed by a stimulus that remains on screen until a response is made. Stimuli are either positive or negative ‘attributes’; in-group or out-group ‘targets’, and must be classified as quickly as possible by making a left (L) or right (R) key press. Reaction times (RTs) are compared between ‘congruent’ blocks, in which positive attributes and in-group stimuli require the same key press; and ‘incongruent’ blocks, in which positive attributes and out-group stimuli require the same key press, to produce an IAT score. (B) Behavioural performance on the IAT can be modelled using a drift diffusion model (DDM) in which two self-excitatory but mutually inhibitory neural populations coding for left and right motor outputs, respectively, noisily integrate external sensory evidence until the firing rate of one population reaches a pre-defined decision threshold. The time taken to reach the decision threshold produces an RT, while the winning population corresponds to the decision made (which may or may not correspond to the sensory evidence presented, i.e. be either correct or incorrect). (C) In our simulations, the sensory evidence provided to each DDM motor population in each IAT trial is determined by activity in the self-image network. Neurons coding for the IAT stimulus – Positive or Negative attributes and in- or out- group targets (Light- and Dark- skinned faces, respectively, in the example shown here) – receive a set level of external sensory input, while additional input to either motor response population arises from recurrent excitation within the self-image network (indicated by thicker coloured arrows). Note that connections from the self-image network to the DDM are flexibly reconfigured between congruent and incongruent blocks. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

Empirical and Simulated Performance on the IAT. (A) Comparison of the IAT effect in empirical data from a population of sixty light-skinned females (see Peck et al., 2013 for details) and simulated data from a population of sixty light-skinned female agents, each of which shows significantly positive implicit bias towards other light-skinned agents (both p < 0.001). (B) Changes in the IAT effect (ΔIAT = IAT_post − IAT_pre) produced by a short VR experience in real participants and simulated agents embodied in a light-skinned (EL); dark-skinned (ED); or alien (i.e. purple skinned) virtual body (EA); or who passively view another dark-skinned agent in the VR environment (NE). In both real and simulated data, embodiment in a dark-skinned virtual body generates a reduction in the IAT effect. Moreover, in the empirical data, embodiment in a light-skinned virtual body generates a small but non-significant increase in the IAT effect, which reaches significance in the simulated data. (C) The measure of bodily resonance produced by the self-image network is modulated by self-esteem: specifically, agents with lower self-esteem (i.e. reduced activity in neurons coding for positive features when the agent perceives its own body) exhibit lower overall bodily resonance (left), and lower firing rates in neurons coding for positive features (right), when observing other agents whose features overlap with their own. This reduction in bodily resonance corresponds to both (D) lower IAT scores and (E) less significant changes in IAT scores after being embodied in a dark skinned virtual body for agents with lower self-esteem. See 2.2, 2.4 for further details. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)