Hallucinations and Strong Priors

Abstract

Hallucinations – perceptions in the absence of objectively identifiable stimuli – illustrate the constructive nature of perception. Here we highlight the role of prior beliefs as a critical elicitor of hallucinations. Recent empirical work from independent laboratories shows strong, overly precise, priors can engender hallucinations in healthy subjects, and that individuals who hallucinate in the real world are more susceptible to these laboratory phenomena. We consider these observations in light of work demonstrating apparently weak, or imprecise, priors in psychosis. Appreciating the interactions within and between hierarchies of inference can reconcile this apparent disconnect. Data from neural networks, human behavior and neuroimaging support this contention. This work underlines the continuum from normal to aberrant perception, encouraging a more empathic approach to clinical hallucinations.

Hallucinations and Perception

“Instead of saying that an hallucination is a false exterior percept, one should say that the external percept is a true hallucination.” [1]

Hallucinations (see Glossary) are percepts without corresponding external stimuli [2]. They characterize many serious mental illnesses like schizophrenia and post-traumatic stress disorder [3]. They occur in the context of Alzheimer’s and Parkinson’s diseases and epilepsy [4], hearing loss [5] and eye disease [6]. But they frequently occur in the absence of any detectable illness, as isolated experiences in up to 50% of people (following bereavement, for example) and in between 2 and 10% of the population on a daily basis [7]. They occur in all sensory modalities, though auditory and visual are most commonly reported. There has been a long and growing appreciation of the constructive nature of perception [1]: it is more than mere receipt of sensory information, instead involving a synthetic process, based upon prior expectancies (henceforth priors) [8]. Whilst this is efficient with regards to information processing [9], it can be prone to error. In particular, this renders perception similar to hallucination [1] [10]. Here we are concerned with the role of priors in generating hallucinations [11].

We focus on auditory verbal hallucinations (AVHs) or voices, though the principles that we outline do, we believe, extend beyond audition. AVHs occur in around 80% of patients diagnosed with schizophrenia but they present us with profound clinical problems: since they are so common, their status as diagnostic markers for mental illness is very uncertain [12]. Moreover, when treatment is indicated, they can, in a significant proportion of people, prove persistent [13]. Thus, there is a real need to understand AVHs. We review work using formal computational models of perception to understand the basis of AVH [14] and suggest that hallucinations arise when prior beliefs exert an inordinate influence on perception [11] [15].

According to predictive coding accounts, perception involves adopting a hypothesis that explains what is causing our current sensation [16], an inference that is optimized by prior knowledge about probable candidates [16]. Those priors are compared to incoming sensation and prediction errors are computed. If priors are more precise than sensory inputs, we consider them strong. They will dominate inference and prediction errors will be ignored [17–20]. On the other hand, relatively more precise prediction errors will dominate priors and drive belief updating (changing ones’ priors for subsequent inference). The powerful contribution of expectation to perception leads us to speculate that hallucinations might arise when prior predictions exert an inordinate influence over perceptual inferences, creating percepts with no corresponding stimuli at all [11] [15].

Data consistent with the strong prior account

That prior experiences could engender hallucinations in healthy volunteers was first reported by Carl Seashore, working at the Yale Psychological Laboratory in 1895. Seashore reported hallucinations in one modality (touch, sight, taste) conditional on learned cues in another sensory modality, and engendered hallucinations by mere suggestion [21]. A more recent example of hallucination-by-suggestion: When healthy volunteers were instructed they would hear the song White Christmas, but instead they were played white noise, 5% reported experiencing Bing Crosby’s voice [22]. People who hear hallucinatory voices in their everyday lives are more susceptible to this effect [23].

Pavlov’s assertion that such inferences were similar to the conditioned reflex [24] has garnered empirical support. Ellson in the 1940s embraced the conditioning paradigm and employed more controlled procedures for consistent stimulus delivery. The illumination of a tungsten filament bulb was presented as a predictor of the presentation of a near threshold 1kHz tone. After learning this association, the tone could be omitted and participants would continue to report tones, conditional on the illumination of the bulb [25]. These studies underline the role of learning in these hallucinations. The predicting stimulus needn’t be visual and the hallucination needn’t be auditory. Tones can likewise be established as predictors of visual stimuli (e.g. dimly illuminated triangles), such that a 1khz tone presentation can engender a visual hallucination (of triangles that were never presented). These experiences even transfer out of the laboratory; subjects report seeing triangles on their television screen, conditional on hearing a 1khz tone, even when no triangles were presented [26].

More recently, visual-auditory conditioning has been employed to demonstrate that voice-hearing patients are significantly more susceptible to this effect than patients without hallucinations and controls [27]. However, these observations (and those that preceded them) could have been driven by the demand characteristics perceived from the experimental situation. More contemporary research has sought to mitigate this concern. Prior knowledge of a visual scene confers an advantage in recognizing a degraded version of that image [28]. Patients at risk for psychosis—and, by extension, voice-hearing—were particularly susceptible to this advantage and its magnitude correlated with hallucination-like percepts. Similarly, there is a version of this effect in audition; voice-hearing participants appear to have an enhanced prior for speech in degraded auditory stimuli even when not explicitly instructed. Indeed, they were able to detect that degraded speech earlier than control participants [29]. It is difficult to attribute these findings to demand characteristics.

A bias toward top-down information appears to be the basis for “sensory conditioning” [27, 30–33], wherein modern psychophysics was brought to bear on the Ellson paradigm, establishing a visual stimulus as a predictor of a difficult-to-detect, near-threshold, auditory stimulus. Participants begin to report auditory stimuli that were not presented on the basis of the visual cue. This effect is amplified in individuals who hallucinate. We recently showed that this effect is mediated by precise prior beliefs, that those priors are more precise in people who hallucinate, and that people with a diagnosed psychotic illness are less likely to update those prior beliefs in light of new evidence [34]. Critically, the neural circuit underlying these conditioned phenomena—including superior temporal gyrus and insula—largely overlapped with the circuit engaged when patients report hearing voices in the scanner [34, 35].

Individuals with more striatal dopamine (itself a marker for psychosis risk [36]) are more susceptible to this impact of prior expectation on auditory perception: those with higher dopamine are more likely to perceive target auditory stimuli embedded in a stream consistent with their expectations rather than what was actually presented to them [37]. Visual hallucinations in Parkinson’s disease have likewise been related to precise priors [38], which may be driven by dopaminergic, cholinergic, and serotonergic perturbations[39]. Whilst we focus here on AVH, we suggest that precise priors in may underwrite hallucinations in other modalities and illnesses. The challenge will be characterizing them and discerning why each specific underlying pathology culminates in voices rather than visions, hallucinations rather than delusions (see Box 1).

Box 1. Halucinations & Delusions – commonalities, differences.

Delusions and hallucinations tend to co-occur though not always. In this computational psychiatry literature, they have been dissociated. Whereas hallucinations have been consistently related to more precise priors, delusions have been associated with less precise priors.

For instance, examining the relation between priors and delusional experiences with ambiguous visual motion stimuli reveals that delusion-proneness both in healthy individuals as well as in schizophrenia patients was related to weaker statistical learning of new priors [111–113]. However, delusion-proneness both in healthy individuals as well as in schizophrenia patients is correlated with high-precision priors induced by a high-level cognitive manipulation (a suggestion that 3D glasses will shift their perception of the ambiguous motion [111, 114]). These apparently contradictory observations might be reconciled by assuming a hierarchy in the brain’s predictive machinery: Here, imprecise perceptual priors (or failure to attenuate sensory precision) may lead to the formation of delusional beliefs. At the same time, precise cognitive priors at a higher hierarchical level will sculpt perception, subtending hallucinations and the maintenance of delusions (by engendering delusion consistent percepts [115]).

This comorbidity of hallucinations and delusions is consistent with the factor structure of psychotic symptoms; a 3-factor solution: Negative symptoms, Disorganization and Positive Symptoms (comprising hallucinations and delusions) [116]. However, there may be structure within the positive symptom factor [117]. Hallucinations may co-occur more frequently with delusions of parasitosis (that there are insects on the skin) and less frequently with grandiose ideas, religious and referential delusions [117]. Might there be psychological and biological evidence in favour of such relationships between symptoms?

A recent study probed priors for the gist vs. the details of ambiguous images in a healthy population with varying degrees of hallucination- and delusion-proneness [118]. Hallucination-proneness

The precision-weighting of priors and prediction errors may be independent at different hierarchical levels and in different modalities [18]. Neurobiologically, this may be mediated by the higher density of recurrent connections in higher-level association cortices, compared with primary sensory regions, such that a psychotogenic perturbation that impacts excitatory/inhibitory (E/I) balance may have more profound effects higher rather than lower in the hierarchy [120]. See [121] for a detailed exposition of the role of E/I balance in learning, inference and psychosis). In brief, the E/I relationships may implement exactly the precision weighting mechanisms that underlie predictive coding. Blocking NMDA receptors (with ketamine for example) profoundly alters E/I balance [122, 123], thus altering the balance between priors and prediction errors [115], perhaps differently at different hierarchical levels [120]. NMDA blockade would decrease the precision of priors facilitating delusion-like inferences [124]. However, ketamine does not typically engender hallucinations (except perhaps in circumstances of high environmental uncertainty[125]). Hallucinations (and sustained delusions) would entail neuroplastic learned changes in priors which may arise from dopamine driven learning mechanisms or serotonergic effects on deep pyramidal cells[115]. These effects are not characteristic of acute ketamine. But they may arise with chronic administration of ketamine or amphetamine or the administration of serotonergic hallucinogens[115]. Comparing and contrasting the different phenomenologies of these drugs[115] as well as psychotic symptoms across different illnesses will be a key test of our hypothesis. We suspect that in each case, where there are hallucinations, there will be high-precision priors.

Taken together, it appears that across sensory modalities prior experiences can induce hallucinatory experiences even in healthy participants. The circuits and neurotransmitters involved appear to align with those implicated in psychotic illness and indeed, people with hallucinations in their everyday lives are more susceptible to these prior-driven effects in the laboratory. The source of those priors and their relationship to the precision of sensory evidence (i.e., prediction errors) in that modality or other modalities (including sensations related to one’s own actions) is critical to understand. We attempt to do so in what follows.

Data that challenge the strong prior account

There are at least two lines of empirical work that challenge our account (i) the apparent immunity to visual illusions and (ii) the failure of corollary discharge in schizophrenia. (Corollary discharges are predictions of the sensory consequences of our actions and, are thought by some, to underwrite our inference about agency for events [40] - a type of prior [41]). Both are considered indicative of less precise priors. We suggest that these data may actually be consistent with the strong prior account – if we allow that these precise priors arise as a mollifying response to the uncertainty engendered by less precise priors elsewhere in the system: either at lower hierarchical levels – in the case of illusions, or in parallel hierarchies in a centrifugal network, as in corollary discharge.

Visual Illusions

Illusory perception departs from actual sensation. For example, artists can create illusory depth on a canvas by mimicking the images that would be formed on the retina by a truly three-dimensional environment [42]. Illusions are thought to be mediated by top-down priors [43]. That is, our prior belief about some feature of sensation sculpts our ultimate percept so it conforms to the belief, rather than the sensation. There are empirical data that demonstrate a reduced sensitivity to some illusions in patients with schizophrenia [44, 45], which tends to suggest that their priors are less precise. However, such effects rarely correlate with symptom severity and when they do, they do not relate to hallucinations specifically. Furthermore, other illusions are enhanced in people with schizophrenia (for example, the three flash illusion – in a manner that correlates with positive symptoms [46]). Hierarchy may be crucial here – since the three flash illusion involves temporal expectations, it is likely to engage more complex and, we would argue, higher-level processing than other more static illusions [47].

It is possible that illusions could fail at low levels of processing and hallucinations could be generated as a response, higher in the hierarchy, as suggested by the work with deep neural networks (see below). In Figure 1, for example, the precision of priors and prediction errors at the sensory level are relatively balanced. When illusions are not perceived by patients with schizophrenia, it could be that they fail to attenuate sensory precision, enabling prediction errors to ascend the hierarchy to induce belief updating. We propose that these un-attenuated (or un-explained away) prediction errors induce a particular sort of high-level prior belief that is the hallucination. There are even data suggesting that psychotic individuals with and without hallucinations utilize different priors to different extents in the same task. People with hallucinations have strong/precise perceptual priors that are not present in psychotic patients who did not hallucinate and who, indeed, may have weak/imprecise priors [34].

Figure 1. An Inferential Hierarchy.

Here we sketch the hierarchical message passing thought to underlie predictive coding and expand on the details of between- and within-layer computations. Sensory input is conveyed via ascending prediction errors in superficial pyramidal cells (red arrows, posited to be mediated by fast AMPA glutamate receptor signaling). Posterior expectations are encoded by the activity of deep pyramidal cells. These cells then provide descending priors (black arrows, mediated by slower and more diffuse NMDA and GABA signaling) that inform prediction errors at the lower level – instantiating the computations that remove the expected input, leaving prediction errors to be assimilated or accommodated, depending on their precision (depicted by the balance, hypothesized to be implemented via slower neuromodulators like dopamine, acetylcholine and serotonin, depending on the particular inferential hierarchy). Lateral interactions (horizontal arrows) mediated within-layer predictions about the precision of priors (black) and prediction errors (red).

Corollary Discharge

The suggestion that a disturbance of corollary discharge, leading to erroneous attribution of inner speech to an external source, can explain voice-hearing [48–53], also appears incompatible with our account. Critically, the theory is that this inappropriate external attribution arises from a failure of prediction of the sensory consequences of one’s own actions, which appears to contradict the idea that hallucinations arise from overly-precise priors.

Corollary discharges are predictions of the sensory consequences of our actions. Normally, they cancel out the expected consequences of actions. Unexpected consequences are more readily attributed to external causes: the more prediction error is generated, the more likely an action (or perhaps a thought) has external (rather than internal) causes [51]. These predictions in the form of corollary discharges have been increasingly referred to as priors in the literature. Their failure in people with schizophrenia has been associated with the severity of psychotic symptoms, though not always hallucinations specifically [48–53]. These data trouble our model because they seem to represent instances of less precise priors causing psychotic symptoms. How can this be?

Much like prediction errors from lower in the hierarchy can be suppressed by updating or increasing the precision of higher level priors, we propose that a similar compensation may occur. According to predictive processing theory, for any given sensory experience, there is a cause to be hypothesized. This cause may be agentic or non-agentic in origin (if non-agentic, the problem is simply one of perceptual identification). If agentic, there is an additional problem: was it self or an external agent? Here, the winning hypothesis will be determined by the relative precision of our predictions about internal and external cues. We suggest that the unattenuated prediction errors arising from the poorly predicted consequences of self-generated actions portend a misattribution of agency. This then leads to an appeal to hypotheses or beliefs that these prediction errors were generated extrinsically: “I am not sure it is me so it must be you”.

Egocentric processing, in the form of corollary discharges, has been detected in multiple brain regions [40]. Its impairment in schizophrenia and psychosis in particular has been demonstrated in eye-movement [54] and force matching tasks [55]– in each case, impaired corollary discharge correlates with the severity of positive symptoms. Some inferences are more impacted by egocentric processing than others [56]. Indeed, visual processing, whilst integrated, possesses dissociable streams for inferences about exteroceptive, interoceptive, and proprioceptive states (i.e. actions). Likewise, speech has reception and motor areas [57]. The auditory cortex receives a range of inputs from these regions, within and between hemispheres, and across different timescales, mediating specific speech predictions and thus specific contents [58], and we would argue agency attributions for perceived speech. It would be interesting to map impairments in each specific predictive mechanism to the contents of particular hallucinations.

Intuitively, imagine that I failed to generate an accurate prediction of my inner speech. This would induce large amplitude prediction errors over the consequences of that speech. Now, these large amplitude prediction errors themselves provide evidence that it could not have been me speaking; otherwise, they would have been resolved by my accurate predictions. Therefore, the only plausible hypothesis that accounts for these prediction errors is that someone else was speaking. Note that this formulation is consistent with both sides of the story. On the one hand, a failure to suppress (covert or overt) sensory prediction errors in this setting can be due to a failure of sensory attenuation - or due to a failure to generate accurate predictions (i.e., corollary discharge). The consequence of these failures is the false inference that I was not the agent of these sensations. Crucially however, this conclusion rests upon precise prior beliefs about sensory precision or the amplitude of prediction errors; namely, “large amplitude prediction errors can only be generated by things I can’t predict”. It is worth noting here that such beliefs mean that prediction errors are informative not just about one’s degree of surprise but also about the origin or nature of an input: perceptions generated by external sources are more surprising than those generated by oneself and perceptions generated by external agents are more unpredictable than those that are non-agentic (for a fuller discussion of this see [59]). Technically, these are known as hyperpriors—in other words, prior beliefs about precision as opposed to content. In short, one can have both an abnormality of sensory precision and prior precision that, together, underwrite the Bayes optimal but false inference that someone else caused my sensations.

Bayesian models of multi-sensory integration emphasize —and empirical work confirms — that the prioritization of streams of information during inference depends on their relative reliability or precision [60–63]. This case has also been made for reinforcement learning [64]. The rubber-hand illusion [65] — to which people with psychosis are more sensitive [66] — is likewise explained as Bayesian integration of visual and proprioceptive sources of information under the prior assumption of a common source [67]. People tend to infer that common cause and thus conclude that a prop hand belongs to their body when they see the hand stroked in synchrony with what they feel (they report feeling ownership of the hand, they show fear responses when the prop hand is threatened [68] and they mount an immune response to that actual hand [68] ), suggesting a malleability of egocentric agency mechanisms. People with psychosis even experience the illusion in the asynchronous control condition [66], perhaps due to an expanded binding window of touch and vision [69], under the influence of a stronger prior on the precision of exteroceptive/allocentric sources relative to interoceptive/egocentric sources [70].

Audiovisual integration of sensory cues is critically dependent upon the relative precisions of the information to be integrated [62]. Thus, if we assume similar integration of external and self-related sources of information in overall inference, when corollary discharge fails, precision of self-processing decreases, and as such weighting shifts towards external sources and reliance on external priors should increase. Thus, weak/imprecise and strong/precise prior effects can co-exist. And failures of corollary discharge can, in theory, give rise to the high precision prior expectation that a voice will be heard – as has been observed [29].

Perturbed integration of internal and external cues will also disrupt sense of agency. For example, internal action cues (experience of an intention, sensory feedback) may be imprecise while the external cues (outcome of action for example) may be precise. This would lead some thoughts and actions to be experienced as unintended and thus externally generated [71, 72]. At the same time, the binding of action and outcome may be increased (due to the relative precision of outcome processing), which has been empirically demonstrated in psychosis, ketamine infusion, and psychosis proneness [73–77].

The co-existence of precise priors for external perception and imprecise for self-agency speaks to the way in which hierarchies of inference are organized, compartmentalized and integrated. When we think about more than one hierarchy, say for inferences about internal states, external states and actions [78], there are – as we ascend each hierarchy – more and more points of intersection, such that, at their apex, there is an amodal model that integrates them into a coherent percept [79]. It is our contention that the relative impact of each hierarchy to ongoing experience is proportional to the precision of its predictions, and, when one element is imprecise or noisy, other hierarchies minimize the precision-weighted prediction error.

Thus, both within and between hierarchies, imprecise priors can subtend more prcise ones and the apparent contradiction between datasets and theories of hallucination is reconciled. We predict that corollary discharge deficits, as measured with the force matching task or two-step saccade task [54], should correlate with strong/precise priors in the tasks reviewed presently.

Strong Priors & Inner Speech

If precise priors do drive hallucinations, then the relationship between inner speech and hallucinations still needs to be explained. Why is the theory that when people hear voices they are hearing their own inner speech so compelling? And how do we explain the functional imaging data derived from symptom capture? When people report hearing voices in the scanner, their speech production and reception circuitry, including Wernicke’s area (within the superior temporal gyrus) and Broca’s areas, are engaged [61]. Non-speech perceptual predictions seem to engage those same brain regions, including superior temporal gyrus [80] [34], which is itself tuned to the statistics of human speech [81]. The activity of speech circuitry during hallucinations may reflect the imposition of priors rather than the presence of inner speech per se. Thus, we can explain the consistent engagement of speech (or we would posit prediction) regions during hallucinations and their specific relationship to hallucination severity. Voices seem likely to be the dominant (though not exclusive) content of auditory hallucinations because our auditory apparatus is tuned to (i.e., has precise priors for) the natural statistics of speech. And they may be experienced as agents communicating, because we believe (based on our overwhelming experience of the contiguity between voices and agents) that voices are typically attached to an agent [82]. The higher-level prior that emerges from paranoia (that one is the cynosure of others’ preoccupations, thoughts and actions) provides a ready-made expectation that there will be agents and communication. In other words, at one level, the voice is an expected characteristic of auditory processing and, in addition, the agency/paranoia prior increases the voice prior (since voices typically emanate from agents). The experience of the voice of course reinforces the higher-level prior. Voices are distressing because they are non-consensual and they engage the highest levels of our inferential hierarchy—those levels that contain our narratives about ourselves [83] - but also because they are omniscient, since the content and not just the form is constructed by expectation, it has the potential to touch on all that preoccupies and worries the person.

Basic Preclinical Mechanisms

Expanding the explanation of hallucinations beyond an inner speech focus proffers a number of advantages. First, and most importantly, it allows a more inclusive definition of voice-hearing to incorporate sensed presences, voices without speech, and the rest of the rich range of hallucination phenomenology that characterizes the lived experience of voices [84]. In proposing a mechanism such as this, it reminds us that beliefs are part of the perceptual process. This is crucial because it highlights that we cannot treat a voice as an isolated symptom or aberration but must include a consideration of their corpus of experience and belief. Indeed, it encourages a much more nuanced view of the gradation from aberrant perceptions to aberrant beliefs [85].

At a more reductionist level of analysis, our account suggests the exciting possibility of modeling hallucinations in species other than humans. Whilst simple conditioning processes may be involved in human language development [86], and clearly other species have language-like processes, these are not as developed and intricate as human speech. Animals cannot report their subjective experiences in a manner that we can understand, unlike humans who hear voices. Prior work on animal hallucinations has tended to focus on objective behavioral responses to psychotomimetic drugs, like head twitch responses under serotonergic hallucinogens and stereotypies induced by amphetamines [87]. Despite the neurochemical relevance (and the potential of these behaviors to guide antipsychotic drug development) [87], to call these hallucinations entails an unacceptable degree of anthropomorphism (to us at least).

Our account of hallucinations avoids this. It is grounded in much earlier work in associative learning theory, wherein learned predictive associative links between stimuli can, on presentation of one of the associates, be so precise that the other is evoked and richly experienced, perhaps as a hallucination. This possibility was first acknowledged by Jerzy Konorski in 1967, who alluded to a ‘perceptual hunger’ that gave rise to hallucinations [88], a sentiment echoed more recently by the suggestion that aberrant prediction errors induce a ‘hunger for priors’, a drive to reconcile errors in expectation that may manifest as a bias towards prior experiences and ultimately hallucinations [89]. With regards to animal conditioning and hallucinations, Konorski argued that conditioned motor responses to an external stimulus in a dog could be elicited by predictors of those stimuli and used to infer that those stimuli were ‘hallucinated’ [88]. This line of inquiry was further developed and formalized by Peter Holland and colleagues [90]. When a hungry rat is presented with a tone and subsequently a sweet sugar solution, the rat learns after only a few trials that the tone predicts sugar. By making the rat nauseous in the presence of the tone, it can be demonstrated that the tone evokes a highly realistic, sensory representation of the sugar, which the rat has trouble distinguishing from reality – because he transfers the nausea associated with the tone to the sugar and subsequently consumes less sugar [91]. With extended training, rats stop having these cue-induced hallucinations [91], but not in animal models that recapitulate the biology of psychosis [92, 93]. Recent work suggests that these representation-mediated phenomena are dopaminergically mediated. For example, behavioral over-expectations based on the rats’ associatively learned model of reward contingencies are driven by dopamine signaling [94–96].

When human participants’ sense of control over their environment is intentionally decreased (with spurious feedback), they tend toward illusory pattern perception, seeing nonexistent signal in noise and detecting illusory trends in the stock market [97]. This uncertainty may also be signaled in part by dopamine [98]. As described above, people who weighted their priors more strongly during perceptual inference have higher levels of striatal dopamine including those who hallucinate outside of the laboratory [37].

The deep hierarchical neural networks described below may be uniquely suited to modeling these representation-mediated over-expectation phenomena [99], since they are capable of capturing the underlying complex, contextually sensitive, associative relationships that drive these effects. We argue the same may be true of hallucinations. The debate continues in artificial intelligence as to whether there should be any innate prescribed structures within deep neural networks [100]. With regards to our account of hallucinations, we do not believe that humans are born with their models of the world pre-configured; we, like neuro-constructivists [101], believe that humans infer their parameter values through experience. But, those values can be innately constrained. For example, the expectation that a caregiver would protect us may be relatively hardwired. If that expectation is violated, we may develop a world model—and a set of social expectations—that color our perceptual inferences in a maladaptive manner [102].

Neural Network Models of Hallucinations

In some of the earliest computational psychiatry, the late Ralph Hoffman implemented an artificial neural network model (a Hopfield network) that could learn outputs given particular input patterns [103]. By pruning the allowable connections Hoffman made the network ‘hallucinate’ spurious outputs [103]. State of the art AI, ‘Deep Learning’ [104] involves stacks of Hopfield-like networks, separated by hidden layers, that learn representations of data, each stage in the hierarchy learns to generate or reconstruct the activation patterns in the stage below. One such-network, the Deep Boltzmann machine (DBM), utilizes both feedforward and feedback processing [6]. Each hidden layer receives input from below, and from above (a kind of prior). Its units learn latent variable representations of the input data. Thus, it can synthesize input data even in the absence of such data [105], which makes DBMs ideal to study hallucinations. Indeed, a DBM to model the genesis of visual hallucinations in Charles Bonnet Syndrome [106], an illness characterized by vision loss accompanied by visual hallucinations. They showed that, in response to low-level impairments, homeostatic mechanisms stabilize network activity levels, such that hallucinations arise when input is lacking, consistent with the observation that de-afferented cortex becomes hyper excitable – we argue this reflects the imposition of explanatory priors on noisy inputs. They assayed the role of hierarchical structure. A DBM trained without the topmost layer failed to learn a generative model and thus failed to show complex hallucinations; hence, no high-level priors, no hallucinations (see Figure 2). They speculate that the balance between feedforward and feedback, between priors and prediction errors, is mediated by acetylcholine [107–109]. Intriguingly, administering scopolamine, a muscarinic cholinergic antagonist, increases conditioned hallucinations in healthy volunteers [31].

Figure 2. How Hierarchies Might Hallucinate.

Problems lower in a hierarchy may generate simple hallucinations. Problems that pervade higher levels will engender more complex experiences. Consistent with the empirical data from humans viewing images generated from deep-networks perturbed at low and high levels [135]and DBM models of perception where in hallucinations only arise under the influence of the highest-level generative model [106]. Figure adapted from [135], licensed under a Creative Commons 4.0 International License (http://creativecommons.org/licenses/by/4.0/.)

These synthetic examples of aberrant perceptual inference fit comfortably with the predictive coding formulation - in the sense that most psychedelics act upon serotonergic neuromodulatory receptors in deep pyramidal cells. It is these cells that are thought to encode Bayesian representations that generate top-down predictions. Having said this, models of perceptual hallucinosis and false inference may not be apt to explain the hallucinations seen in psychosis. This brings us back to the distinction between perceptual and active inference. Our explanation for hallucinations above rests explicitly on beliefs about agency and how “I cause sensory consequences”. In short, to develop a formal understanding of the sorts of hallucinations in conditions like schizophrenia, one might suppose that deep priors about how we actively generate our sensorium lie at the heart of auditory hallucinations, which are generally comorbid with delusions of control and other false concepts (see Box 1). This has the important implication that subsequent studies of hallucination may focus not so much on prior beliefs about the exteroceptive world, but on how we infer the consequences of our action across all (exteroceptive, interoceptive and proprioceptive) domains. This is particularly prescient given the engagement of the anterior insula and other frontal regions in hallucinations.

These computational models provide much needed support and constraint in what we mean when we evoke hierarchical layers of processing. It is clear that modulating the function of high and low levels of the hierarchy have different consequences for the presence and content of hallucinations in network models. We observe similar hierarchical organization of beliefs relevant to hallucinations in computational analyses of human behavior [34]. In neuroimaging data, high-level predictions about multi-sensory learned associations correlated with activity in supramodal regions such as orbitofrontal cortex and hippocampus [110], whilst dynamic low-level predictions were associated with activity in primary cortices [110]. Both levels were implicated in hallucinatory perception, but clinical AVH were associated with dysfunction in higher-level neural and behavioral responses [34]. The explanatory power of hierarchical differences – rather than being a salve for data that challenge the theory - appears to have empirical reality.

Concluding Remarks

Arguing whether hallucinations are driven by high- or low-precision prior beliefs may seem arcane. However, we believe it has profound clinical implications. If priors are dominant, then the neurobiological [34] and psychological [102] interventions indicated would be different from if they are imprecise. At the very least, the contemporary data suggest a precision-weighted trade-off for the contents of experience, in higher versus lower hierarchical layers and more or less agentic hierarchies, which mediate the different contents and complexities of hallucinations. More broadly, the observation from our laboratories, that people who do not hallucinate in their everyday lives can nevertheless evince hallucinations in the laboratory, favors the active inference model of perception and agency. We hope that this understanding of hallucinations as an exaggeration of normal non-hallucinatory perception, to which we are all sometimes prone, may help further our understanding of perception and cognition (see Box 2) and go some way to de-stigmatize clinical hallucinations and encourage empathy for those who experience them.

Box 2. Cognitive penetration.

Many cognitive scientists have argued that cognition does not influence perception [126], that perception and cognition are separate modules. We believe the hallucinations data outlined presently are inconsistent with such encapsulated modularity. This is because learned beliefs about the task — ones that were not present within the perceptual module prior to the experiment— can be conditioned and appear to change perception.

In The Modularity of Mind (1983), Jerry Fodor sketched a blueprint of mental architecture comprised of modules—systems that process a single specific kind of information [127], strictly segregated into discrete mental faculties that can be damaged in isolation [128]. An encapsulated perceptual system, kept separate from the influence of beliefs, could have the advantage of keeping our beliefs grounded in the truth offered by our senses [129]. However, a cognitively penetrable perceptual apparatus may be equally adaptive, despite misperceiving, as long as the resulting behavior is adaptive [130, 131].

Staunch modularists might allege that the top-down effects of cognition on perception are merely an effect of attention [132]. This seems hard to reconcile with the new percepts we appear to be able to condition. Furthermore, they may allow for the type of learning we describe by suggesting that new priors that are learned in our tasks are compiled into the perceptual module with experience [133]. This seems like an influence of cognition on perception to us. Predictive perception need not necessarily demand penetration [134], but penetration of perception by cognition, exemplified by hallucinations, is at least consistent with hierarchical versions of predictive processing.

Outstanding Questions.

• Are corollary discharge and strong prior effects anti-correlated, as our account would predict?

• Is the sensitivity to conditioned hallucinations in people who hallucinate domain specific? That is, are people with AVH also more susceptible to visual conditioned hallucinations?

• What are the neurochemical mechanisms of strong priors? Can they be remediated with pharmacological manipulations?

• Are the contemporary neural data a cause, a consequence or a correlate of hallucinations? We will need causal manipulations, like transcranial magnetic stimulation and effective connectivity analyses to make causal and directional claims about the neurobiology of priors and hallucinations.

• Does this account apply to hallucinations trans-diagnostically (i.e. outside of the context of schizophreniform illness, to hallucinations in Alzheimer’s, Parkinson’s, Charles Bonnet Syndrome, head injury and stroke)?

Highlights.

• Recent data establish a role for strong prior beliefs in the genesis of hallucinations. These data are difficult to reconcile with aberrant inner-speech theories, in which weaker predictions about the potential consequences of ones’ own inner-speech drive an inference that speech is emanating from an agent external to oneself.

• In the predictive-coding view, this failure of self-prediction renders the consequences of one’s inner speech surprising. The prediction errors induced are explained away by the strong higher-level priors identified in recent work. The presence and contents of hallucinations can be understood in terms of learning, inference and a reliability-based trade-off between internal and external information sources, biased toward high-level priors.

• If priors divorce perception from sensation somewhat, then the distinction between hallucination and perception becomes less clear. We hope this explanation renders hallucinations more understandable and less stigmatizing.

GLOSSARY

Agency

The inference that one has acted to effect some particular result

Allocentric

Having ones’ interest and attention directed externally (often toward other persons)

Charles-Bonnet Syndrome

Named after a Swiss Philosopher, who wrote about his Grandfather’s experiences losing his sight to cataracts. Common amongst those who have lost their sight, this condition causes visual hallucinations. The sight loss can also result from macular degeneration, glaucoma or diabetic eye disease

Corollary Discharge

The command sent to the muscles to produce a movement, directed to other regions of the brain to inform them of an impending movement

Deep-Boltzman Machine

A type of generative neural network (see below) that can learn internal representations

Delusions

Fixed, false beliefs that persist despite overwhelming contradictory evidence. Often held with considerable preoccupation and distress

Egocentric

Having one’s interest and attention directed toward the self (and away from others)

Exteroceptive

Relating to stimuli that are external to the organism

Hallucinations

Percepts without corresponding external stimulus

Illusions

Something likely to be mis-perceived by the sensory apparatus. Typically, an erroneous conclusion about something actually present

Interoceptive

Relating to stimuli that are internal to the organism

Neural Network

A computing system vaguely inspired by nervous systems, based on a collection of connected units or nodes which loosely model neurons

Precision

A property of priors and prediction errors (see below). Formalized as the inverse variance or negative entropy. Precision underwrites notions like certainty, confidence, salience and reliability

Prediction Error

The mismatch of expectancy and experience that can, if precise (see above), lead to belief updating

Predictive Coding

A scheme used in signal processing that compresses signals using a predictive model. Predictive processing models of brain function suggest that similar processing efficiencies are employed by the brain. This entails a generative model of expected inputs and a means to broadcast (and learn from) failures in those predictions (prediction errors – see above)

Prior

Ones’ beliefs about some quantity or dimension before new evidence is taken into account

Psychosis

Experiences that depart from consensual reality

Voices

Or Auditory Verbal Hallucinations, experiences that someone is talking when no-one is around