Theory of mind: a neural prediction problem

Abstract

Predictive coding posits that neural systems make forward-looking predictions about incoming information. Neural signals contain information not about the currently perceived stimulus, but about the difference between the observed and the predicted stimulus. We propose to extend the predictive coding framework from high-level sensory processing to the more abstract domain of theory of mind; that is, to inferences about others' goals, thoughts, and personalities. We review evidence that, across brain regions, neural responses to depictions of human behavior, from biological motion to trait descriptions, exhibit a key signature of predictive coding: reduced activity to predictable stimuli. We discuss how future experiments could distinguish predictive coding from alternative explanations of this response profile. This framework may provide an important new window on the neural computations underlying theory of mind.

Figure 1. A Sensory-Coding-Based Model of Social Predictive Coding

Predictor neurons (P) code expectations about the identity of incoming input and pass down the prediction to lower level predictor neurons (green arrows) and lower level error neurons (blue arrows). Error neurons (E) act as gated comparators, comparing sensory input from lower levels (red arrows) with the information from predictor neurons (blue arrows). The difference between the predicted input and the actual input is passed up to higher level error neurons, propagating up the processing hierarchy (red arrows). Error neurons also modulate the response of predictor neurons (purple arrows), likely both by inhibiting the predictor neurons making incorrect predictions, and enhancing predictor neurons making correct predictions. When the information that is being passed up from lower levels matches the information carried by the predictor neurons, the error neurons’ response to the input is reduced, “explaining away” the predictable input (Rao and Ballard, 1999).

Figure 2. Three Brain Regions Involved in Different Aspects of Theory of Mind: Examples of Individual Regions of Interest

A region in posterior superior temporal sulcus (STS, green, peak voxel [66, −36, 12]) involved in action perception (localized using biomotion relative to scrambled biological motion, Pelphrey et al., 2003); a region in temporo-parietal junction (TPJ, blue, peak voxel [62, −52, 18]) involved in thinking about beliefs and desires (localized using stories about mental states relative to stories about physical events, Saxe and Kanwisher, 2003), and a region in medial prefrontal cortex (MPFC, red, peak voxel [−2, 56, −4]) involved in thinking about people’s stable preferences and personalities (localized using attribution of traits to self relative to attribution of traits to someone else, Mitchell et al., 2006). All three ROIs localized using single subject data, p < 0.001.

Figure 3. Predicted Stimuli Elicit Reduced Activity with Sharpened Representations

Kok et al. (2012a) report that an accurate prediction led to reduced response amplitude in early visual cortex, but also simultaneously to an “improved” stimulus representation, as measured by multi-voxel pattern analysis. Consistent with this suggestion, we find that in the right temporo-parietal junction (RTPJ; right box) the response amplitude to a predicted belief was lower, but the spatial pattern associated with that belief category was more reliable. Left: a sample stimulus. All stories described first a harmful action, and then the agent’s belief. The “predicted” belief (solid arrow) was consistent with the action (i.e., making the act an intentional harm). The “unpredicted” belief (dotted arrow) was inconsistent and rendered the harm an accident. Middle: The amplitude of response in the TPJ was lower for the intentional than accidental condition. Right: The spatial pattern of response in the TPJ was most robust and reliable across trials for intentional harms, and somewhat less reliable for accidental harms. Data from Koster-Hale et al. (2013).