If the predictive coding framework is correct, it provides a new view not only on memory function but also on how we think about memory systems and the relation between memory and other cognitive systems. This framework provides a theory about the role of simulation models in the relationship between reflexive forms of self-consciousness and the narrative self (Hohwy 2007). It provides a theoretical explanation of the finding that memory systems are also involved in perception[8] and interoception. This implies that we not only simulate offline (e.g., mental time travel, mind-wandering), but also simulate online. The simulated model provides us with a subjective reality through which we see the external world and ourselves. It is transparent and immediate: We experience it as objectively real and we directly interact with what is represented.
However, this characteristic is absent in patients suffering from depersonalization. Depersonalization is an example of how one can become detached from one’s simulated model of oneself: One’s mental autobiography is no longer direct, and one experiences a sense of distance from the model.[9] Gerrans (this collection) suggests that the loss of sense of presence in depersonalized patients results from a failure to minimize prediction error from the hypoactivity of the AIC—the activation of which informs us of the significance of external or internal information. Gerrans’ theory is based on Seth et al.’s idea of interoceptive inference (or interoceptive predictive coding; see also Seth this collection), according to which predictive coding not only applies to exteroception but also to interoception, and emotional states, including the sense of presence, arise from interoceptive prediction successfully matched to actual interoceptive signals (Seth 2013; Seth et al. 2011). As it is suggested that the AIC is suggested to be the correlate of the integration of exteroceptive and interoceptive signals and that it plays a role in maintaining a salience network for the relevant states, the hypoactivity of the AIC leads to the failure to associate affective significance with bodily states. As Gerrans suggests, “not all higher level control systems can and do smoothly cancel prediction errors generated at lower levels” (this collection, p. 9). Because the coding formats at each level are distinct, the coding format of low-level processing is opaque to introspection (p. 9). The problems faced by depersonalized patients can be accounted for by the prediction error based on persisting, unexpected hypoactivity. Attention is then directed towards resolving the prediction error. Gerrans’ proposal is that an inability to explain away the surprisal and this increased attention causes anxiety in DPD. Here, CD can be seen as a strategy for some systems to react to anxiety in order to minimize the prediction error.
As Gerrans suggests, “[d]elusions are best conceptualized as higher-level responses to prediction error which, however, cannot cancel those errors” (this collection, p. 10). That is, even though not all prediction error can be successfully cancelled, the brain—the organ that constantly minimizes prediction error, according to predictive coding framework—still tries to modify its model in order to decrease surprisal, though unsuccessfully. If what I have suggested in the last section is correct, the function of memory systems is to update knowledge contributing to the construction of predictive models in order to minimize prediction error. The anomalous model of CD is thus one constructed by the hierarchical simulation model to match the hypoactivity of the AIC—the loss of appraisal that represents the significance of self-related information. To construct a model in which oneself is dead or does not exist cannot successfully explain away the prediction error—since one still has the experience of a bodily state—it may nevertheless be the best solution the given system can come up with in order to cope with the increased anxiety resulting from increased attention.
However, this still leaves us with the question of why some depersonalized patients develop CD, whereas most of them do not develop this delusion. Gerrans (2014) suggests that the difference between delusional and non-delusional minds lies in differences in the default mode network, which include information that triggers activity, hyperactivity, and hyperconnectivity, interaction with the salience system, and absent or impaired “decontextualized supervision” (pp. 73–74). Decontextualized supervision allows one to “reason about oneself using impersonal, objective rules of inference” (p. 76).[10] The activity of its circuit is anti-correlated with the activity of the default mode network (pp. 83–84) because of the limited cognitive resources for high-level metacognitive processes. Gerrans suggests that delusional thoughts arise from the system’s failure to balance this allocation; thus they slip through the supervision system.
Nevertheless, the existence of decontextualized supervision explains how anomalous forms of predictive models—which would be suppressed in non-delusional subjects—could emerge, but it does not account for the model’s relation to anomalous experience or to the way in which the content of delusion is constructed (e.g., Cotard delusion). I therefore propose that a delusional mind does not only result from a compromised decontextualized supervision; it also results from an aberrant precision expectation[11] of exteroceptive or interoceptive signals. Jakob Hohwy (2013) proposes the notion of uncertainty expectations: We predict the causal structure of the world (and of one’s own bodily state), as well as the level of uncertainty in the environment, which allows us to respond to the external environment under various levels of uncertainty. The strength of prediction error is proportional to the expected certainty: When the uncertainty level is expected to be higher (due to external or internal noise), the prior model is weighted higher, whereas expected low uncertainty gives more weight to bottom-up prediction error. According to Hohwy (2013), delusion arises when precision expectation is either too high or too low, and those in between would report only the anomalous experience, without forming a delusion. In the case of Cotard delusion developed from depersonalization, when one has the expectation of high precision, the system tends to be driven by the bottom-up predictive error of unexpected hypoactivity of the AIC, rather than the prior model. One is, therefore, more likely to revise the model in order to explain away the surprisal resulting from the mismatch between the actual and predicted activation level of the AIC; that is, the systems of patients suffering from CD are driven by an urge to modify their top-down predictive models in order to conform to the loss of AIC activity. The construction of the model in CD is considered an attempt to minimize prediction error.
Finally, explaining delusion under the predictive coding framework provides new understanding to the debate between one- and two-stage models of delusion. The one-stage model holds that anomalous experience only suffices to explain the occurrence of delusion (Gerrans 2002; Maher 1974, 1988); according to two-stage model, however, other cognitive disruption is required to explain the content of the delusion in particular (Young & De Pauw 2002). However, if the predictive coding framework is correct, the clear distinction between experience and rationalization assumed in the traditional discussion does not exist: Perception, cognition, and action are now considered continuous and highly integrated (Clark 2013b; Hohwy & Rajan 2012). Experience and rationalization are different layers of abstraction within the very same process of prediction error minimization under the predictive coding framework.