The mind is organized as a hierarchical system that uses representations of the world and its own states to control behavior. According to recently influential Bayesian theories of the mind, all levels of the cognitive hierarchy exploit the same principle: error correction (Friston 2003; Hohwy et al. 2008; Jones & Love 2011; Clark 2012, 2013; Hohwy 2013). Each cognitive system uses models of its domain to predict its future informational states, given actions performed by the organism. When those predictions are satisfied, the model is reinforced; when they are not, the model is revised or updated, and new predictions are generated to govern the process of error correction. Discrepancy between actual and predicted information state is called surprisal and represented in the form of an error signal. That signal is referred to a higher-level supervisory system, which has access to a larger database of potential solutions, to generate an instruction whose execution will cancel the error and minimize surprisal (Friston 2003; Hohwy et al. 2008). The process iterates until error signals are cancelled by suitable action.
This is a very basic outline of the predictive coding idea dodges a crucial question: the extent to which Bayesian formalisations actually describe neurocomputational processes rather than serving as a predictive calculus for neuroscience (Jones & Love 2011; Hohwy 2013; Clark 2012; Park & Friston 2013; Moutoussis et al. 2014). It also blurs an important distinction which is not salient to formalisations such as Bayesian theory: namely the fact that not all higher level control systems can and do smoothly cancel prediction errors generated at lower levels. For example vision and motor control are good examples of predictive coding systems (Hohwy 2013). Often however experiences best explained as carrying information about prediction error are not cancelled by the adoption of a higher-level belief. Consider déjà vu experiences which signal mismatch between an affect of familiarity and perception of a novel scene (O’Connor & Moulin 2010). We know the scene is novel, but it still feels familiar. The point is just that the higher order belief does not always smoothly cancel prediction error. And this should be expected. Coding formats are not uniform across cognitive systems, which is why sensory and higher-level cognitive integration is such a cognitive achievement for the mind.
From our point of view what matters are the key ideas of hierarchical organization, upward referral of surprisal and top-down cancellation of error. Also crucial is the idea that the highest levels of cognitive control involve active, relatively unconstrained, exploration of solution space. This is the level at which attention can be redirected to alternative solutions and their imaginative rehearsal. Phenomena such as delusion represent a high level response to an obstinate signal of prediction error that cannot be simply cancelled from the top down. This way of thinking of the mind weds a version of predictive coding theory to insights from neurocomputational theory that treat executive systems as specialized for the resolution of problems which cannot be solved at lower levels. Thus at low levels in the hierarchy the structure of priors and errors and referral of surprisal is constrained, modularized some might say. At the so-called personal level of belief fixation predictive coding best describes the idea that those experiences which command executive resources are those which signal prediction error which cannot be resolved at lower perceptual and quasi perceptual levels. This is at least one level at which predictive coding involves active sampling of information (active inference) as well as the routine cancelling of surprisal according to a well defined prior model. The latter almost defines perception. The former, according to O’Reilly & Munakata 2000) as well as predictive coding theorists (Spratling 2008) is definitive of executive control.
Thus most of the detection and correction of error occurs at low levels in the processing hierarchy at temporal thresholds and using coding formats that are opaque to introspection. Keeping one’s balance, parsing sentences and recognizing faces are examples. We have no introspective access to the cognitive operations involved and are aware only of the outputs. This is the sense in which our mental life is tacit: automatic, hard to verbalize, and experienced as fleeting sensations that vanish quickly in the flux of experience. This is the “Unbearable Automaticity of Being” (Bargh & Chartrand 1999). However even these relatively automatic processes generate experiences of which we can become aware. The recognition of faces, for example, produces an affective response within a few hundred milliseconds. When that affective response is absent or suppressed due to malfunction a prediction is violated and the discrepancy between familiar face and lack of familiar affect is referred to higher levels of executive control to deal with the problem.
At the higher levels of cognitive control, surprisal is signalled as experience that becomes the target of executive processes. These metacognitive processes evolved to enable humans to reflect and deliberate to control their behaviour. The highest levels of cognitive control involve reflection, deliberation, rehearsal and evaluation of alternative courses of action and explicit reasoning. When for example a predicted affect is absent we might find ourselves in the position of a patient described by Brighetti who lost affective responses to her family and her professor. She had “identity recognition of familiar faces, associated with a lack of SCR [SCR is skin conductance response, a measure of electrodermal activity consequent on affective processing]” (Brighetti et al. 2007). In other words her predicted affective response to familiars was absent, which resulted in an experience becoming the target of higher-level control processes. Such patients sometimes produce the Capgras delusion that the familiar person has been replaced by an imposter or double. A truly florid delusion such as is sometimes seen in schizophrenia might elaborate the delusional thought into an epic paranoid narrative.
The aim here is not to enter into the controversy about the explanation of the Capgras delusion but to note the role of the architecture that generates it (Young et al. 1994; Breen et al. 2001; Ellis & Lewis 2001). Higher levels of cognitive control are engaged to deal with error signals referred from lower levels in the hierarchy. Perhaps the most important level in the hierarchy for personal and social life is the level at which subjectively adequate narratives are generated to make experience intelligible and by which we communicate our experiences to others. This is the level at which delusional thoughts originate. By subjectively adequate here I merely mean “fits the experience of the subject”. At even higher levels of cognitive control we can revise and reject those subjectively adequate autobiographical narratives, replacing them with empirical theories that draw on publicly available norms of reasoning and semantic knowledge to produce objectively adequate responses to subjective experience (Gerrans 2014). Delusions are best conceptualized as higher-level responses to prediction error which, however, cannot cancel those errors. In fact as Clark (2013) points out such delusory models in effect “predict” further experiences of that type, which means that the delusion will be strengthened.
A very important point to note for the subsequent explanation of depersonalization and the Cotard delusion is that it is not the absence of affect per se which produces the error signal and engages higher-level cognition. Lack of affective response alone does not require a high level response unless that lack of affect is unpredicted. That is why we are not bothered by lack of response to strangers (we don’t predict it at any level in the control hierarchy) but if a new mother has no affective response to her baby the experience can be part of a syndrome of post-natal depression.
The example of post-natal depression allows us to make another important point about the relationship between predicted affect and psychosis. Mothers most vulnerable to post-natal depression are those who had powerful positive expectations of motherhood and the bond with the infant. When that bond does not materialize for some reason they are confronted with a distressing lack of predicted affective response. Sometimes this will produce a kind of Capgras delusion regarding the baby. The mother might say that the baby has been replaced or is an alien (Brockington & Kumar 1982). Interestingly, and tellingly, if the mother is also extremely anxious the condition can be even more serious. Anxious attention to the experience tends to magnify the problem.
This role for anxiety is nicely elucidated by the predictive coding framework. Formal considerations aside, the concept of predictive coding places a huge emphasis on the signaling of error. This means that incoming information must be compared to a prediction and the difference computed and referred to a control system. At higher levels those error signals take the form of experiences. These experiences are often imprecise and opaque since they are produced by lower level systems that encode information in different formats to those used by explicit metarepresentational capacities. They also compete for metarepresentational resources among the constant flux of experiences that engage attention. Thus they create a problem of working out for any experience how much is signal and how much is noise.
It is very important for high-level cognition to be targeted as precisely as possible for only as long as required. Thus any vagueness in experience needs to be resolved. Attention is the process which solves this problem. Hohwy (2012, p. 1; my emphasis) makes the point for perceptual inference but it applies in general:
conscious perception can be seen as the upshot of prediction error minimization and attention as the optimization of precision expectations during such perceptual inference.
Clark (2013, p. 190) makes a similar point:
Attention, if this is correct, is simply one means by which certain error-unit responses are given increased weight, hence becoming more apt to drive learning and plasticity, and to engage compensatory action.
The point is that attention is directed to error signals in order to make them more precise by increasing the signal to noise ratio. Attention amplifies the signal and maintains it while higher-level systems try and interpret the experience and manage appropriate responses. If the response works the error signal is cancelled and attention can be directed elsewhere.
Within this framework we can make an observation about anxiety that can be overlooked by approaches that concentrate on the arousal, hypervigiliance or the associated beliefs concerning threat or danger. These approaches de-emphasise a crucial element. That is uncertainty. Anxiety is an adaptive mechanism that primes the organism cognitively and physiologically to resolve uncertainty. Thus, if a prediction cannot be verified, or an error signal disambiguated, anxiety in this sense will result. Of course what we call pathological anxiety is the dysfunctional activation and maintenance of these mechanisms. The point is that someone who is anxious in this way will continue to misallocate attentional, cognitive and physiological resources to experiences. Another point about anxiety is that, in pathological cases, action does not cancel the signal or the dysfunctional allocation of resources to it. This may be why the role of anxiety in depersonalisation is not straightforward. Some recent studies have not found a strong correlation between anxiety and depersonalisation (e.g., Medford 2012). However the scales used to measure anxiety give a score that sums scores for self-report of feelings, behaviour and cognition. The suggestion here is that what really matters is the allocation of attention to signals which cannot be resolved, perhaps because they are intrinsically noisy, ambiguous or have insufficient information. It is also important that the patient cannot resolve the uncertainty by revising the predictive model that generates it since that is usually maintained low in the predictive hierarchy by mechanisms that are not accessible. The person with Capgras delusion, for example, automatically predicts affective response to familiar faces and when it goes missing there is nothing she can do to revise that prediction. Instead she is confronted with an anomalous experience, which automatically captures attention. Similarly with depression. Loss of affective response is not something that can be restored from the top down.
In some cases of post-natal depression all these factors seem to be operative. The mother expected to bond with the infant but in fact perhaps birth was traumatic, the baby did not attach straightaway, and the mother needed more support and reassurance than she received. She was left distressed and unable to cope which made bonding and attachment even more difficult. This would be bad enough but if the mother had a strong prior expectation that motherhood would be straightforwardly rewarding a prediction is violated. If the mother is also anxious she will attend intensively to the resultant experience of absent affect, but she will encounter only further feelings of emptiness and panic. The presence of the baby and the expectations of family and friend only compound the sense that she is not feeling what she should be feeling. What happens next depends on context and support but it is not really surprising, especially given the relationship between massive hormonal fluctuation and emotional regulation, that in some cases new mothers develop psychotic symptoms (Spinelli 2009).