71. Computational Psychiatry and the Bayesian Brain

Once the “new kid on the block,” computational psychiatry has quickly grown into a substantial field, with dedicated annual conference (https://www.cpconf.org/), workshops, courses, and its own specialist journal (https://cpsyjournal.org). So, what is computational psychiatry? In brief, it is the use of formal models of brain function to characterize the mechanisms of psychopathology—usually in a way that can be articulated in computational or mathematical terms. Perhaps the more interesting question is why computational psychiatry? It is well known as editorial folklore that prepending “computational” to any discipline or journal name immediately reduces its impact factor by at least two points. So, why take the risk? There are two compelling answers. First, it establishes a focus for winning the “hearts and minds” of colleagues from a variety of areas, including funding agencies and next-generation psychiatrists. Second, computational psychiatry enforces a focus on the mechanisms, quantification, and formal understanding of psychopathology that is essential for scientific enquiry and more informed (i.e., mechanistically grounded) psychiatric phenotyping or nosology.

The aim of this chapter is to provide tangible examples of computational psychiatry and how it motivates research in neuroscience that can be translated into the clinical sciences. In what follows, we first consider the properties a computational formulation must possess to be useful in psychiatry. We then focus on a particular example, namely active (Bayesian) inference and predictive coding. The second section highlights the importance of this formulation by showing how it contextualizes other formal approaches. We conclude with selected examples of how theoretical principles reveal mechanistic themes that unify apparently disparate aspects of psychiatric disorders. We touch on functional dissociative symptoms, soft neurological signs in schizophrenia, interoceptive inference and autism, dysconnection models of delusional beliefs, and formal models of interpersonal exchange. These examples are chosen to illustrate the breadth of psychopathology that can be subsumed under a formal treatment—and how they can all be understood in terms of a singular pathology: the false inference that can be ascribed to neuromodulatory failures at the synaptic level.

There are many formal or computational schemes that could be called on to characterize psychopathology, ranging from parallel distributed processing or neural network theory and dynamical systems theory to reinforcement learning and game theory. However, these theoretical frameworks do not address the quintessential nature of the most severe functional deficits in psychiatry: the production of false beliefs. The things that occupy psychiatrists are, almost universally, attended by abnormal beliefs and their behavioral sequelae, such as dysmorphophobia, paranoid ideation, delusion, hopelessness, lack of self-worth, suicidal ideation, obsession, disorientation, false memories, and so on. This calls for computational frameworks that rest explicitly on inference or beliefs and their neurophysiological realization.

Fortunately, recent years have seen a paradigm shift in cognitive neuroscience that offers exactly the right sort of computational theory that allows one to talk about false beliefs—and understand how these arise from pathophysiology at the synaptic level (Figure 71.1). This shift is away from the brain as a passive filter of sensations—or an elaborate stimulus-response link—toward a view of the brain as a statistical organ that generates hypotheses or fantasies to be tested against sensory evidence. In short, the brain is now considered a fantastic organ (fantastic: from Greek phantastikos, the “ability to create mental images,” from phantazesthai) (Friston, 2013). For many people, this perspective can be traced back to Helmholtz and the notion of unconscious inference (Helmholtz, 1866/1962); however, in the past decade or so, the basic idea has been formalized and generalized to cover deep or hierarchical Bayesian inference—about the causes of our sensations—and how these inferences induce beliefs, movement, and behavior.

Modern formulations of Helmholtz’s notion are now among the most popular explanations for information processing in the brain and are usually considered under the Bayesian brain hypothesis through the lens of predictive coding (Rao & Ballard, 1999; Srinivasan et al., 1982). Predictive coding is not a normative or descriptive scheme—it is a process theory with a biologically plausible backstory. There is now considerable anatomical and physiological evidence supporting the notion of predictive coding in the brain (e.g., see the work of Bastos et al. [2012] for a review of canonical microcircuits and hierarchical predictive coding in perception, and consult Adams, Shipp, et al. [2013] for corresponding treatments of the motor system).

In these schemes, neural representations in higher levels of cortical hierarchies generate predictions of representations in lower levels. These top-down predictions are compared with representations at the lower level to form a prediction error (i.e., hypothesized to be associated with the activity of superficial pyramidal cells). This mismatch signal is passed back up the hierarchy to update higher representations (i.e., hypothesized to be associated with the activity of deep pyramidal cells). This recursive exchange of signals suppresses prediction error at each level to provide a hierarchical explanation for sensory inputs that are predicted at the lowest (sensory) level. In computational terms, neural activity is thought to encode beliefs or probability distributions over states in the world that cause sensations (e.g., my visual sensations are caused by a face that is smiling). The simplest encoding corresponds to representing the belief with the expected value of a hidden cause or expectation. These causes are referred to as hidden because they have to be inferred from their sensory consequences.

In summary, predictive coding represents a biologically plausible scheme for updating beliefs about the world using sensory samples (Figure 71.2). In this setting, the brain—and the entirety of computation and biology it entails—can be regarded as a distillation of causal structure in the world that embody a generative model for creating predictions of sensations in a Helmholtzian sense. The implications for perception are illustrated well in Kandel’s (2012) discussion of “the beholder’s share” (see also Figure 71.3):

The insight that the beholder’s perception involves a top-down inference convinced Gombrich that there is no “innocent eye”: that is, all visual perception is based on classifying concepts and interpreting visual information. One cannot perceive that which one cannot classify. (p. 204)

If our brain embodies a generative model of our world, then much of it must be occupied by modeling other people. In other words, we spend most of our time predicting the proprioceptive and exteroceptive consequences of behavior (both mine and yours). To fully appreciate the bilateral nature of these predictions, one needs to consider inference in an embodied context. In brief, perception can be understood as resolving exteroceptive prediction errors by selecting predictions that best explain sensations, whereas behavior suppresses proprioceptive prediction error by changing sensations. This suppression rests on classical reflexes, whose equilibrium points are supplied by descending proprioceptive predictions (Adams, Shipp, et al., 2013). This is active inference and simply involves equipping a predictive coding scheme with reflexes (Figure 71.2). At a more theoretical level, it also brings the Bayesian brain hypothesis into the inactive domain—in the form of control and planning as inference (Botvinick & Toussaint, 2012).

Clearly, if high-level sensorimotor expectations provide top-down predictions of the sensory consequences of moving, then they also provide a repertoire of hypotheses for inferring the intentions of others. This is because the exteroceptive (e.g., visual) consequences of your movements can be predicted from the consequences of making the same purposeful movement myself—all I need to do is to infer the agency (self vs. other). This may be important for understanding false beliefs about agency in schizophrenia (Frith et al., 2000) and provides a nice explanation for mirror neurons that respond to self-made acts and during action observation (Kilner et al., 2007).

However, to harness the mirror neuron system during action observation, we must attenuate proprioceptive prediction errors that would otherwise elicit movements in our own body that mirror the subject of our observation (cf. echopraxia). This attenuation rests on reducing the precision of proprioceptive prediction errors. Precision can be regarded as a measure of signal to noise or confidence assigned to an information stream. Mathematically, precision is the inverse variance or reliability of a signal. Estimating precision speaks to a fundamental aspect of inference in the brain, namely the encoding of precision or expected uncertainty (Iglesias et al., 2013; Yu & Dayan, 2005). In other words, we must infer not only the content of our sensorium but also the context, in terms of its (expected or subjective) precision. This represents a subtle but generic problem that the brain must solve, where the solution may rest on modulating the gain or excitability of neuronal populations reporting prediction error (Friston, 2010).

Heuristically, one can regard ascending prediction errors in cortical hierarchies as broadcasting “newsworthy” information that has yet to be explained by descending predictions. However, the brain also has to select the channels it listens to. It can do this by adjusting the volume or gain of prediction errors that compete to update expectations. There is empirical evidence (Iglesias et al., 2013) that this precision weighting of prediction errors is a central computational process throughout the brain and may be instantiated through neuromodulatory mechanisms of gain control at a synaptic level. This neuromodulatory gain control corresponds to a Bayes-optimal encoding of precision in terms of the excitability of neuronal populations reporting prediction errors (Friston, 2010). This casts computational light on why superficial pyramidal cells are replete with synaptic gain control mechanisms, such as N-methyl-d-aspartate (NMDA) receptors and classical neuromodulatory receptors such as D₁ dopamine receptors (Doya, 2008). Furthermore, it places excitation–inhibition balance in a prime position to mediate “precision-engineered” message passing within and among hierarchical levels (Humphries et al., 2009).

This contextual aspect of predictive coding has been associated with attentional gain control in sensory processing (Feldman & Friston, 2010) and has been discussed in terms of affordance in the context of active inference and action selection (Cisek, 2007). Crucially, the delicate balance of precision over hierarchical levels has a profound effect on veridical inference—and may hold the key for a formal understanding of false beliefs in psychopathology (Adams, Stephan, et al., 2013). See Powers et al. (2017) for a treatment of hallucinosis in this setting.

Key challenges for formal accounts of brain function are emotion, self-awareness, and their disorders. Recently, a framework called interoceptive inference has been introduced to provide a computational account for these affective constructs. As it is typically described (Gu et al., 2019; Seth & Friston, 2016), interoceptive inference makes two related but dissociable claims. The first claim is that approximate Bayesian inference about physiological states of the body underlies feeling states. This entails an important claim that “feelings” are a certain kind of belief state that are updated during inference to provide the best explanation for interoceptive sensations (e.g., “I am anxious” is the best explanation for current bodily sensations). The second claim is that physiological states, generating interoceptive sensations, are controlled using active inference. In other words, autonomic reflexes work to align internal states with descending predictions from an agent’s generative model of themself. This is precisely analogous to active inference as a tool for controlling the motor system (Adams, Shipp, et al., 2013). Taken together, interoceptive inference, with these two major claims, reconciles the conflict between the James–Lange and the Canon–Bard theories in the sense that they are both right—that subjective feelings are both cause and consequence of autonomic states (for a detailed discussion, see Gu et al., 2019).

The foregoing treatment may seem a bit too monothematic to cover all that computational neuroscience has to offer psychiatry. However, part of the construct validity of active inference is that it contextualizes other formal approaches. For example, formal models of schizophrenia are often cast in terms of neuronal disconnection, which makes them amenable to study with parallel distributed processing schemes. There are two versions of disconnection hypothesis: the first implied by Wernicke’s sejunction hypothesis, which postulates an anatomical disruption or disconnection of association fibers (e.g., Hoffman, 1997), and the other that postulates abnormalities at the level of synaptic efficacy and plasticity, leading to dysfunctional integration or dysconnectivity among cortical and subcortical systems (Stephan et al., 2009). One can see immediately that dysfunctional integration at the synaptic level coincides exactly with aberrant (neuromodulatory) precision control in predictive coding—and earlier theories framed in terms of reduced signal to noise (Braver et al., 1999). Furthermore, this putative abnormality fits comfortably with nearly every synaptic or physiological theory of schizophrenia, ranging from dopaminergic and NMDA receptor dysfunction (Lisman et al., 2008) to GABAergic (γ-aminobutyric acid–ergic) abnormalities (Gonzalez-Burgos & Lewis, 2012) and dysfunctional excitation–inhibition balance (Jardri & Deneve, 2013). The common theme here could be a failure to maintain the appropriate gain of principal or pyramidal cells (Barch et al., 2012).

Another example of computational psychiatry is the appeal to reinforcement learning and optimal decision theory. These formal frameworks have shaped systems neuroscience over the past decade or so. They provide a normative and heuristic account of choice behavior with clear neurobiological correlates. Perhaps the most celebrated correlate is the association of dopamine with reward prediction error in the context of temporal difference models of reward learning (Schultz et al., 1997). So, how does this square with dopamine as a candidate for encoding precision or uncertainty (Fiorillo et al., 2003)? The answer is straightforward: In active inference, reward and value are treated as prior beliefs that determine what is predicted, including one’s choice behavior. Prior expectations (i.e., preferences) therefore play the role of reward (in reinforcement learning) or utility (in behavioral economics). In this setting, a negative reward prediction error signals a loss of confidence or precision in expectations about rewarding outcomes (Friston, 2013). More recent applications of reinforcement learning to understanding psychopathology are usually framed in terms of Bayesian inference and prior beliefs. Perhaps more tellingly, formal theories of aberrant salience—originally based on notions from reinforcement learning (Kapur, 2003)—are now more naturally cast as theories of aberrant precision, particularly given the formal connection between precision control and attention or salience (Feldman & Friston, 2010). The focus on aberrant precision also opens the door to theories of metacognition prevalent in psychology and social neuroscience. Metacognition, or the study of beliefs about beliefs, often focuses on reporting the confidence in decisions. This affords a measure of insight (a metacognitive dʹ) that has clear relevance for psychiatry and a direct link to subjective certainty or precision (Rouault et al., 2018).

A subtle but important advantage of casting reward or value functions in terms of prior beliefs—and their precision—is that one can now parameterize the beliefs that account for individual choice behavior. This enables a quantitative and formal phenotyping, in terms of beliefs and attitudes. This approach to computational phenotyping lies at the heart of many current computational psychiatry initiatives—and has even been extended to game theoretic models of interpersonal exchange (Moutoussis et al., 2014; Na et al., 2021). These extensions may be particularly important for characterizing various psychopathies and their genetic or physiological correlates (Moutoussis et al., 2014; Na et al., 2022).

Finally, there is growing interest in characterizing neuronal dynamics using concepts from dynamical systems theory. These range from understanding neuronal avalanches in terms of self-organized criticality (Plenz & Thiagarajan, 2007) to characterizing neurophysiological time series in terms of multistability (Deco & Jirsa, 2012). Indeed, changes in the nature and deployment of coherent or coordinated dynamics have often been associated with conditions such as schizophrenia—and have been promoted as evidence of functional dysconnection (Phillips & Silverstein, 2003).

So, how do criticality and itinerant dynamics relate to inference? Simply stated, they provide a rich dynamical repertoire that enables the brain to respond quickly to changing inputs (Deco & Jirsa, 2012). In terms of predictive coding, weakly stable fixed points—which underlie itinerant dynamics—play the role of hypotheses-generating top-down predictions. This means that a rich dynamical repertoire in neuronal states becomes a rich portfolio of hypotheses that can be called on to explain sensory data. In fact, it is relatively easy to show that optimal inference necessarily precludes very stable fixed points—in exactly the same way that Occam’s principle precludes overly detailed explanations.

If we throw into the mix the fact that precision has to be inferred—and that precision has a profound effect on dynamical stability—we have self-organized criticality. It is this delicate control of instability that renders precision a key parameter in neuronal dynamics (cf. temperature in statistical thermodynamics). In the next section, we discuss what would happen if the encoding of subjective precision were compromised.

This section has focused on formal models from systems and cognitive neuroscience—as opposed to normative models from cognitive science and psychology. There is a principled reason for this: Normative models of cognitive phenomena describe what the brain does, whereas process models describe how the brain does it. Our focus on process models reflects the end point we have in mind, namely to link the phenomenology of psychiatric conditions to their neurophysiological and molecular causes. Perhaps a debate about the utility of normative models—relative to process models—should be placed high on the emerging agenda of computational psychiatry.

At this point, we see how easily one gets from a principled computational framework for action and perception to the synaptic mechanisms that might underlie false inference in psychiatric conditions; in brief, the formal constraints implicit in predictive coding require modulatory gain control on ascending prediction errors. An article by Edwards et al. (2012) exemplifies how one can understand functional (hysterical) symptoms as false inference about the causes of abnormal sensations, movements, or their absence. This example offered a simple neurophysiological explanation of symptomatology that is otherwise rather difficult to diagnose or formulate. This theme is emerging repeatedly in psychiatry: from false inference as an account of positive symptoms (i.e., hallucinations and delusions) in schizophrenia (Adams, Stephan, et al., 2013) to the loss of central coherence in autism (Pellicano & Burr, 2012). Moreover, it is remarkable that the same role for precision weighting of prediction errors emerges from different theoretical treatments of learning and inference in the brain—including predictive coding in vision (Rao & Ballard, 1999), free-energy accounts of perception and behavior (Friston et al., 2006), and hierarchical Bayesian models of learning (Iglesias et al., 2013).

This section discusses a few examples of how these principles have been applied to psychiatric conditions, with a special emphasis on common themes and empirical advances that are likely to be seen during the next decade. Recall that one of the principal roles of precision is to contextualize sensory information in relation to prior beliefs. Perhaps one of the most important examples of this is the attenuation of sensory precision, which enables prior beliefs about proprioception to elicit self-made acts.

A recurrent theme in many psychiatric disorders is a failure of sensory attenuation, with secondary consequences for the acquisition and deployment of hierarchically deep models of the world and interpersonal interactions. In the context of low-level exchanges with the world—such as pursuit eye movements—a failure of sensory attenuation means that sensory precision is too high in relation to the precision of higher prior beliefs about the causes of sensations. It is relatively easy to reproduce the cardinal deficits of slow pursuit eye movements in schizophrenia by simply reducing prior precision in simulations of eye tracking using predictive coding and oculomotor reflexes. This provides a formal explanation for the inability of schizophrenic subjects to infer regular high-order contingencies that underlie target movement and anticipate its motion (e.g., as revealed using occluders).

More interestingly, because prior expectations are compromised, they explain the paradoxical improvement in pursuit in the context of violations (e.g., unpredicted changes in target motion). This is because prior expectations are no longer appropriate and confound the tracking behavior of normal subjects. This example makes the more general point that the relative precision of sensory and prior prediction errors is a crucial determinant of our susceptibility to illusions and responses to unpredicted events or their omissions (Adams, Stephan, et al., 2013). Practically, this speaks to the utility of relatively simple, well-established paradigms—such as slow pursuit eye movements, the mismatch negativity paradigm, and the psychophysics of illusions—in psychiatric phenotyping. The idea here is that behavioral and neuronal responses can be parameterized in terms of precision in the setting of hierarchical predictive coding (Iglesias et al., 2013).

Crucially, quantities such as prediction error and precision have clear neurobiological correlates that make them amenable to modeling. For example, if prediction errors are reported by superficial pyramidal cells in the cortex, then they are directly accessible using noninvasive electromagnetic techniques—because it is these cells that contribute most to event-related potentials and induced responses. Similarly, if precision is encoded by the excitability or gain of superficial pyramidal cells, then this gain can be estimated using biophysical modeling of neuronal circuits (dynamic causal modeling) on the basis of evoked electrophysiological responses.

Much current work rests on working at paradigms that elicit prediction errors, while manipulating their precision to enable the quantification of hierarchical message passing in normal subjects and, ultimately, psychiatric cohorts. If successful, these sorts of studies should provide a computationally and biophysically grounded phenotyping of psychiatric dysconnection syndromes. These can then be used in conjunction with psychopharmacological manipulations and genetic studies to tie down the precise synaptic mechanisms and their molecular basis. Although these low-level paradigms exploit the formal constraints offered by computational psychiatry, they do not really touch on the deeper beliefs that characterize psychosis. So, do we have formal models of delusions?

It is surprisingly easy to simulate delusional beliefs because hierarchical Bayesian inference schemes such as predictive coding are cast explicitly in terms of posterior and empirical prior expectations. Perhaps the best example to date addresses a key issue in schizophrenia research, namely beliefs about agency. The idea here is that psychotic subjects fail to contextualize the consequences of their actions and make false inferences about the agency or owners of their sensory consequences. This is well illustrated in the peculiar resistance of schizophrenic subjects to the force-matching illusion (Shergill et al., 2005). In brief, normal subjects show sensory attenuation during self-made acts, whereas schizophrenic patients appear to show a failure of sensory attenuation. In terms of predictive coding, the attenuation of sensory precision is necessary for an intended movement to temporarily suspend attention to exteroceptive and proprioceptive cues that report the absence of that movement (e.g., we do not attend to the visual motion induced by our saccadic eye movements). In the force-matching illusion paradigm, sensory attenuation reduces the magnitude of self-produced forces relative to externally generated sensations. Crucially, schizophrenic patients are resistant to this illusion and can accurately report the forces that they produce themselves (Shergill et al., 2005). This can be simulated using predictive coding of somatosensory and proprioceptive cues by precluding an attenuation of sensory precision (Brown et al., 2013). However, this comes at a price: To produce the self-generated force in the first place, it is necessary to increase prior precision so that the prior belief that one is moving overrides the sensory evidence that one is not. The problem here is that to explain the precise sensory information—that the force is always less than predicted—the agent must infer an opposing external force. This is a good example of a simulated delusional belief that rests on one simple manipulation—a failure to attenuate sensory precision and compensatory increases in precision at higher levels of the hierarchy (Figure 71.4).

The implication is that one might expect to see abnormalities in the gain control of pyramidal cells in cortical hierarchies. These reports are starting to appear. For example, Fogelson et al. (2014) used event-related potentials and dynamic causal modeling to show “the differences between recurrent inhibitory connections during the processing of predictable and unpredictable stimuli were markedly attenuated in schizophrenics” (p. 204). Similarly, dynamic causal modeling of functional magnetic resonance imaging signals suggests a selective reduction in recurrent inhibitory connections within the medial prefrontal cortex (Bastos-Leite et al., 2015). Subsequent detailed modeling of synaptic efficacy and effective connectivity has started to drill down on the specific populations—and intrinsic connections—that may be responsible for the loss of gain control in schizophrenia (e.g., Adams et al., 2022). The hope here is that the biophysical models used to characterize task evoked and spontaneous activities will become increasingly constrained in functional (computational) terms. So far, we have focused on the computational architectures of inference and false beliefs—but what does this perspective have to say about neurodevelopmental psychopathology?

Perhaps the best example of computational approaches to neurodevelopmental syndromes is found in research on autism spectrum disorder (“autism” hereafter). Recently, much of the phenomenology of autism has been cast in terms of false Bayesian inference that results from a loss of prior precision, relative to sensory precision (Pellicano & Burr, 2012). However, in autism, the consequences of increases in (or a failure to attenuate) sensory precision are being interpreted in a developmental context—in which one needs to consider the consequences for acquisition or learning of deep hierarchical models. This is particularly interesting in relation to interoceptive inference because it touches on the acquisition of generative models that distinguish between self and other. One line of thinking here is that a failure to contextualize interoceptive cues, elicited by interactions with the mother, precludes a proper attribution of the agency (self vs. [m]other) to the interoceptive consequences of affiliative interactions. This hypothesis has several interesting implications for attachment, theory of mind, and a lack of central coherence that characterizes the disorder in later life. It also provides an interesting explanation for interoceptive hypersensitivity (cf. an emotional echopraxia) in autism and failure to engage with prosocial (exteroceptive) cues (Gu et al., 2015). If this explanation is correct, then it provides a clear pointer to abnormalities of (precision) gain control in cortical systems mediating interoceptive inference, such as the anterior insular and cingulate cortex, and transactions with others. Formal theories of autism, as outlined, highlight the importance of inference about others (i.e., social inference), which is a potentially important facet of computational psychiatry in its own right.

Optimal Bayesian decision theory (or game theory) provides a potentially important construct to quantify beliefs—as expressed in interpersonal exchange. Indeed, game theory has already proved useful in characterizing autism (e.g., Yoshida et al., 2010). Although important for computational phenotyping, game theory is a descriptive (normative) framework that does not directly speak to mechanisms. However, by recasting value or utility functions as prior beliefs, one can convert game theory (and expected utility) problems into inference problems. This allows one to use the formal constraints on Bayesian belief updating to obtain an idea about how the brain might arrive at optimal decisions or choices in social settings. Note that optimality rests explicitly on my prior beliefs and that what is optimal for you may not be optimal for me. This is important because it provides the latitude to characterize perception, choices, and interpersonal behavior in terms of subjective priors and ask whether these differ systematically among different personality traits or psychiatric groups. In the context of precision, this line of thinking provides added impetus to quantifying confidence and beliefs in schizophrenia and other disorders—an endeavor that calls on a number of relatively easy paradigms, such as the “rock paper scissors” and “beads” tasks, which reveal a tendency for schizophrenics to “jump to conclusions” (Moutoussis et al., 2011).

Technically, modeling games in terms of active inference requires a move away from predictive coding in continuous time to formulate things in discrete time. The ensuing message passing (based on variational Bayes) has some remarkable similarities to message passing in the brain. In particular, it highlights the recurrent nature of message passing and the pivotal role of precision in selecting the messages that are passed. Of particular interest here is the putative role of dopamine in encoding the precision of beliefs about desired outcomes (Friston et al., 2012). This suggests that economic games can be used to formally characterize behavior in terms of prior beliefs about outcomes and our confidence in those beliefs. See the work of Moutoussis et al. (2014) for a provisional discussion of these ideas in the context of beliefs about others—for example: How do my prior beliefs about the sort of person you are affect posterior beliefs about myself and my subsequent interactions with you?

The examples were chosen to show how a formal approach can provide generic explanations for psychopathology that are physiologically grounded. However, one might ask, Is the notion of aberrant precision (i.e., neuromodulation) so inclusive as to be nonspecific? Clearly, the game at hand is to understand the regional specificity and myriad neuromodulatory mechanisms (e.g., different neurotransmitters acting on different receptor subtypes in different parts of the brain) in relation to functional anatomy and neurodevelopment. It may be that this understanding will be necessary to understand the diversity of psychiatric conditions—and the mechanistic basis of their nosology.

We have seen how computational psychiatry calls on formal models of perceptual inference and learning to provide a mechanistic and functional perspective on psychopathology and its underlying pathophysiology. This review has focused on inference as the overarching theoretical framework, largely because this is the natural way to formalize perception and behavior in terms of probabilistic beliefs. By assuming the brain engages in some form of active inference, one can associate neural dynamics and message passing with Bayesian belief updating and make some remarkably specific predictions about the impact of functional or synaptic dysconnections. We have focused on neuromodulatory failures and how they can be understood in terms of an aberrant encoding of subjective precision or uncertainty—leading to false inference that can be expressed at many different levels. This computational approach necessarily enforces a mechanistic and quantitative view of psychopathology—a view that can accommodate phenomenology ranging from soft neurological signs in schizophrenia to theory of mind in autism, using exactly the same computational principles. One might anticipate that the traditional psychiatric nosology will be enriched by the quantitative and parametric characterizations offered by computational psychiatry. Furthermore, the use of formal models may disclose those levels of description that may, or may not, be appropriate for particular psychiatric conditions.

KJF is funded by the Wellcome Trust. XG is funded by the National Institute of Health (National Institute of Mental Health and National Institute on Drug Abuse), the Wellcome Trust (Leap), and The Simons Foundation (SFARI). This chapter is based on two original articles published in Lancet Psychiatry and Psychopharmacology.