From hallucinations to synaesthesia: a circular inference account of unimodal and multimodal erroneous percepts in clinical and drug-induced psychosis

This theory-building pre-print (2021) provides a mechanistic explanation of how psychedelic and psychotic hallucinations are enacted through ascending and descending circular message-passing networks. Psychedelic phenomena are proposedly related to the amplification descending top-down predictions that over-integrate sensory information into multimodal hallucinations, whereas psychotic states entail the amplification of bottom-up sensory information that becomes overinterpreted as delusions and unimodal hallucinations.

Authors

Back, S. E.
Bouttier, V.
Denève, S.

Published

May 13, 2021

Psyarxiv

meta Study

Links

Read Paper DOI Google Scholar

Abstract

Psychedelics are known to distort perception and induce visual and multimodal hallucinations as well as synaesthesia. This is in contradiction with the high prevalence of distressing voices in schizophrenia. Here we introduce a unifying account of unimodal and multimodal erroneous percepts based on circular inference. We show that amplification of top-down predictions (descending loops) leads to an excessive reliance on priors and aberrant levels of integration of the sensory representations, resulting in crossmodal percepts and stronger illusions. By contrast, amplification of bottom-up information (ascending loops) results in overinterpretation of unreliable sensory inputs and high levels of segregation between sensory modalities, bringing about unimodal hallucinations and reduced vulnerability to illusions. We delineate a canonical microcircuit in which layer-specific inhibition controls the propagation of information across hierarchical levels: inhibitory interneurons in the deep layers exert control over priors, removing descending loops. Conversely, inhibition in the supragranular layers counterbalances the effects of the ascending loops. Overall, we put forward a multiscale and transnosographic account of psychosis with important theoretical, conceptual and clinical implications.

Unlocked with Blossom Pro

Research Summary of 'From hallucinations to synaesthesia: a circular inference account of unimodal and multimodal erroneous percepts in clinical and drug-induced psychosis'

Introduction

Hallucinations are percepts experienced in the absence of corresponding external stimulation and occur across clinical, neurological and drug-induced contexts. Classic serotonergic psychedelics (for example LSD, psilocybin, DMT/ayahuasca) typically produce rich visual and multimodal phenomena including synaesthesia, intensified imagery and greater susceptibility to visual illusions. By contrast, schizophrenia more commonly features unimodal auditory hallucinations (voices) often coupled with reduced sensitivity to illusions. These phenomenological differences coexist with distinct neurochemical accounts—serotonergic agonism for psychedelics and dopaminergic dysregulation (with occasional glutamatergic/GABAergic/serotonergic involvement) for schizophrenia—yet a unified mechanistic explanation spanning molecular, circuit and perceptual levels has been lacking. Leptourgos and colleagues set out to provide a multiscale, unifying account of unimodal and multimodal erroneous percepts using the circular inference (CI) framework. They aim to show, using computational simulations and a proposed laminar microcircuit implementation, how different forms of unchecked recurrent information flow (ascending versus descending loops) can produce distinct phenomenological profiles—crossmodal hallucinations and synaesthesia versus segregated unimodal hallucinations—and how these may map onto different inhibitory mechanisms and neuromodulatory influences in cortex. The study therefore links Bayesian/message-passing models at the meso-scale with laminar inhibitory microcircuits at the micro-scale to account for psychotic experiences across drug-induced and clinical states.

Methods

The investigators formalise perception as hierarchical Bayesian inference implemented via belief propagation (message passing) on generative graphs. They extend prior single-stream models to a two-hierarchy generative model representing two sensory modalities (for example audition and vision) that share a common top node representing a multisensory cause. Nodes are binary (feature present/absent) and messages are iteratively exchanged between connected nodes to compute beliefs. A key methodological innovation is modelling circular inference: insufficient cancellation of redundant messages produces either descending loops (feedback/top-down corrupting feedforward information) or ascending loops (feedforward/sensory signals corrupting priors). The authors also propose a revised algorithmic formulation in which corrective operations occur at the level of beliefs rather than messages; this formulation maps directly onto excitatory–inhibitory interactions in cortex. Simulations used two variants of the model: a symmetrical architecture with equal hierarchy depths for both modalities, and an asymmetrical architecture in which anatomical differences (for example more synaptic relays in the visual stream) create modality-dependent effects. Three simulation scenarios were tested. Scenario 1 (synaesthesia): one modality received strong sensory evidence while the other received evidence against the stimulus, with no prior (LL_P = 0). Scenario 2 (sensory-driven hallucination): both modalities received weak, noisy evidence near chance, again with no prior. Scenario 3 (perceptual illusion / prior-driven hallucination): a strong prior contradicted the sensory evidence. Simulations ran for 30 iterations (each iteration an exchange of messages up and down the hierarchies). Many example parameter settings are reported: in one synaesthesia example LL_A = 2.9 and LL_V = -2; in an illusion example LL_A = LL_V = -1.4 with LL_P = 1. Sensory and prior weights in presented simulations were typically set to 0.95. The authors report having tested robustness across a range of weights, loop strengths, likelihoods and priors. To link algorithmic CI to neurobiology, the researchers derived a neural interpretation of the belief-update equation: beliefs arise from excitatory inputs from levels above and below, balanced by lateral inhibition via interneurons. They propose that interneurons driven by feedforward inputs (putatively in granular/deeper layers, e.g., L4 and/or deep layers) remove descending loops, whereas interneurons driven by feedback inputs (putatively in supragranular layers, e.g., L2/3 and/or L1) remove ascending loops. A simplified canonical cortical microcircuit is outlined in which superficial pyramidal cells integrate information to form beliefs and laminar-specific interneurons implement the inhibitory subtraction operations required to prevent circularity.

Results

Simulations showed qualitatively distinct outcomes depending on whether descending or ascending loops were present and on network symmetry. Across scenarios, the absence of loops produced near-Bayesian (chance or sensory-driven) beliefs; the addition of loops produced strong beliefs unanchored in the immediate sensory evidence, but with different phenomenological signatures. Scenario 1 (synaesthesia): In the symmetrical model, when strong auditory evidence (e.g., LL_A = 2.9) co-occurred with visual evidence against a stimulus (LL_V = -2) and no prior, exact inference produced an auditory percept without a visual percept. Adding descending loops led to cross-amplification: auditory feedback into the visual hierarchy overcame contrary visual evidence and produced a concomitant visual belief (synaesthesia), together with overconfidence in the auditory modality. By contrast, adding ascending loops amplified information within each modality but degraded cross-modal communication, yielding strong unisensory percepts and making synaesthetic inversion unlikely. Scenario 2 (hallucinations): When both modalities received weak, noisy evidence near chance, descending loops generated correlated, crossmodal beliefs—either both modalities believing in the presence of the stimulus or both believing in its absence—producing multisensory hallucinations. Ascending loops yielded segregated, modality-specific amplification: a chance fluctuation in one modality could drive a strong belief there while the other modality held an inconsistent belief, producing unimodal hallucinations and inter-modality inconsistency. Crossmodal hallucinations under ascending loops were possible only when independent unimodal aberrant beliefs co-occurred by chance. Scenario 3 (illusions / prior-driven hallucinations): With a contradicting prior (example LL_A = LL_V = -1.4 and LL_P = 1), descending loops amplified the prior and produced beliefs close to certainty, corresponding to perceptual illusions or strong imagery. Ascending loops produced the opposite effect: stronger reliance on sensory evidence, reduced susceptibility to illusions and weaker imagery. Across simulations, descending loops were associated with increased susceptibility to priors and enhanced integration across modalities, whereas ascending loops promoted reliance on local sensory evidence and segregation of modalities. Additional simulations with the asymmetrical model (visual hierarchy containing more relays and beliefs generated farther from sensors) produced modality-specific predictions: descending loops favoured auditory inputs to dominate visual ones, while ascending loops amplified auditory but not visual processing under certain parameterisations. The authors report that results were robust across a range of weights and loop strengths. They also note a computational signature of descending loops: a generalised over-confidence affecting both modalities, which differs from alternative models that would predict asymmetric confidence changes.

Discussion

Leptourgos and colleagues interpret their findings as providing a unifying, multiscale account linking phenomenology (macro-scale), probabilistic message-passing (meso-scale) and laminar cortical circuits (micro-scale). They argue that different types of circular inference produce distinct clinical phenotypes: descending-loop–dominated CI yields over-integrated hierarchies with crossmodal hallucinations, synaesthesia, stronger mental imagery and greater susceptibility to illusions—attributes commonly reported under serotonergic psychedelics—whereas ascending-loop–dominated CI yields segregated, unimodal hallucinations and reduced vulnerability to illusions, which map more closely onto typical schizophrenia phenomenology. The proposed laminar implementation attributes control of descending loops to inhibition driven by bottom-up inputs (interneurons in L4 and possibly deep layers) and control of ascending loops to feedback-driven inhibition (interneurons in L2/3 and/or L1). From this mapping, the authors suggest neuromodulatory dissociations: serotonergic action at 5HT2A receptors (enriched in deep-layer pyramidal cells) could preferentially disrupt the circuit that prevents descending loops, offering a mechanistic link to psychedelic phenomenology; dopaminergic (and cholinergic) abnormalities implicated in schizophrenia could preferentially affect the circuit that limits ascending loops. Several testable predictions are advanced. Clinically, different types of CI could explain heterogeneity in psychosis: most schizophrenia patients may be dominated by ascending-loop impairments (type A), while a minority with multisensory or treatment-resistant hallucinations may have descending-loop dysfunctions (type B) and relatively greater serotonergic involvement; this has potential implications for differential treatment responses (for example, clozapine's serotonergic profile in refractory hallucinations). Experimentally, descending loops should produce overconfidence across modalities, a signature distinguishable from alternative models, and recent advances in high-field laminar fMRI could allow empirical tests of the laminar, input-specific inhibitory predictions. The authors acknowledge limitations. The proposed microcircuit is simplified and omits interneuron subtypes and within-layer details. The computational account focuses on drug-induced synaesthesia and is not designed to explain developmental synaesthesia, though conceptual parallels exist. They emphasise the need for better clinical characterisation of multimodal hallucinations and more empirical work to link CI signatures to human neurophysiology and animal models. Overall, the paper presents a transnosographic CI-based framework that generates specific, falsifiable predictions and connects algorithmic dysfunctions to laminar inhibitory mechanisms as a route to understanding diverse psychotic perceptual phenomena.

View full paper sections

INTRODUCTION

Hallucinations can be defined as percepts occurring while the person is awake and without corresponding external stimulation of the relevant sensory organ. Although they sometimes occur in non-clinical populations, hallucinations often constitute the hallmark of psychiatric disorders, such as schizophrenia or borderline personality, and are common symptoms in neurodegenerative diseases. Interestingly, hallucinatory experiences can also be induced by psychotomimetic drugs. A particular class of hallucinogenic drugs, known as "classic psychedelics", has fascinated scientists for more than a century. Those include naturally occurring chemicals such as mescaline (extracted from the peyote cactus), psilocybin ("magic mushrooms") and N,N-dimethyltryptamine (DMT), as well as synthetic compounds such as lysergic acid diethylamide (LSD). Long before the first experimental investigations with mescaline, various cultures used the psychoactive properties of these drugs either to improve the physical performance of hunters or to gain spiritual guidance. This is, for instance, the case of the Shipibo shamans who typically drink the ayahuasca brew (which contains DMT) while sitting in a dark place, using songs and perfumes to shape their visions. Interestingly, those Amazonian tribes recognized the capacity of psychedelics to enhance interactions between sensory modalities long before the discovery of LSD-induced synaesthesia. All classic psychedelics are serotonergic agonists with a high affinity for 5HT 2A receptors. These receptors mediate most of the psychoactive effects of psychedelics, as demonstrated by the blocking ability of 2A antagonists (e.g., ketanserin;. Nevertheless, a number of other receptors, including 5HT 2C , 5HT 1A , and 5HT 5A as well as dopaminergic and beta-adrenergic receptors, have also been proposed to play a role in these effects. 5HT 2A receptors are found in both the cortex and subcortical regions but are predominantly expressed in cortical layer V pyramidal cells, suggesting a cardinal involvement of deep cortical layers in the phenomenology of psychedelics. From a neurophysiological point of view, serotonergic drugs increase activity in a variety of cortical regions, including the primary visual cortex and more frontal areas. These serotonergic drugs also cause profound changes in the functional connectivity of the default-mode network and within/between resting-state networks and task-positive networks. Finally, psychedelics can decrease the power of alpha-band oscillations, which has been interpreted as an increased excitability in the absence of external stimulation. At the phenomenological level, psychedelics induce profound changes in people who consume them. Psychedelics notably induce perceptual, emotional and cognitive alterations, while they can also generate mystical experiences and result in a diminished sense of self ("ego-dissolution") and a feeling of unboundedness. Perceptual abnormalities comprise elementary and complex hallucinations (mostly visual or crossmodal) and intensification of visual illusions and mental imagery, together with synaesthesia, an otherwise rare perceptual phenomenon in which activation of one modality leads to subjective experiences in other modalities as well. Interestingly, the content of these hallucinatory experiences (e.g., the "spirits" in the case of the Shipibo shamans) can be modulated by the activation of other sensory modalities ("effect of setting"; e.g., by singing songs or spraying perfumes). The content can also be influenced by the emotional state of the consumer prior the administration of the drug ("effect of set";). In summary, serotonergic hallucinogens generate rich experiences, including a dominant crossmodal component (complex hallucinations with synaesthesia) and a top-down component (increased mental imagery, illusions and emotional effects). This description is very different from the psychotic experiences usually observed in schizophrenia. At the molecular level, schizophrenia has been linked to increased presynaptic storage and release of striatal dopamine. However, glutamatergic, GABAergicand serotonergicabnormalities have also occasionally been associated with these dopaminergic dysregulations. At the phenomenological level, patients with schizophrenia mainly report hearing voices with dominant negative affective content, although a minority of patients also describe multisensory (usually audio-visual) hallucinations. In schizophrenia, hallucinations are regularly found coupled with a reduced sensitivity to illusions. These differences immediately raise new questions: What links exist between serotonergic agonism and the aberrant crossmodal experiences previously described? Are drug-induced psychoses functionally and mechanistically linked to schizophrenia-related psychoses? And if so, what mechanism(s) is(are) at the roots of this phenomenological variability? The recent renaissance of psychedelic science together with the burgeoning field of computational psychiatryrecently brought those questions to light, and a number of insightful theories started to address them. Despite those efforts, a unifying, multiscale account of psychosis ranging from psychedelics to schizophrenia is still lacking. In the first section of this paper, we will integrate available findings in a unique computational framework that is able to capture the different facets of these psychotic experiences. We will notably defend the idea that circular inference (CI), a form of suboptimal hierarchical probabilistic inference in which likelihood and prior corrupt and amplify each other, can offer a holistic and functional explanation for psychoses beyond schizophrenia. Using simulations, we will show how different suboptimal inferences may be linked to various forms of hallucinations. This will allow us to propose a link between observations made at the meso-scale (e.g., erroneous messages passing) and those made at the macro-scale (e.g., behaviour and phenomenology). In a second section, we will review empirical evidence supporting a link between meso-scale and micro-scale findings, in other words, between different forms of CI and the function of inhibitory interneurons in different layers of the cortical column. Our demonstration will build upon the critical role played by the balance between excitatory (E) and inhibitory (I) inputs in information processing within neural circuits. We state that one of the overarching goals of those interneurons is to regulate the neural E/I balance and consequently ensure that the feedforward and feedback flows of information are not redundant.

THE CIRCULAR INFERENCE FRAMEWORK

The brain presents a highly recurrent architecture in which lateral/feedback connections dominate feedforward inputs with a ratio of 9:1. These circuits generate large levels of spontaneous neural activity, directly questioning how the system disentangles self-generated signals from true/new sensory events. This problem seems particularly acute for perceptual inferences, in which sensory cues have to be integrated with prior expectations. Such integration requires both feedforward and feedback connections, incidentally creating internal information loops. According to the CI framework, a finely tuned balance between neural excitation (E) and inhibition (I) in neural circuits could keep the information flow under surveillance, removing all redundant messages. E/I balance is a well-known property of brain circuits. A dysregulation of the E/I balance could be due to impaired inhibition, too much excitation or disruptions in the neuromodulatory systems. The CI framework postulates that if inhibition is insufficient, uncontrolled recurrent excitation results in a reverberation of externally triggered sensory evidence and/or internally generated prior expectations. Such control can be insured by inhibiting redundant messages (those that have already been sent up or down the hierarchy), hence the importance of E/I balance in predicting and cancelling recurrent excitation. Eventually (if this system is overwhelmed), the normally distinct bottom-up and top-down sources of information are corrupted by each other, and the messages are over-counted. A "descending loop" is defined as the corruption of the feedforward sensory information by the feedback (top-down) information, leading to an amplification of the priors. Conversely, an "ascending loop" (also called "climbing loop") is generated when the sensory evidence corrupts the prior, leading to the amplification of the likelihood and an overinterpretation of sensory data. Such circularity could be an important feature of perceptual inference in humans, while in extreme cases, it could generate psychotic symptoms, including hallucinations and delusions. This idea is in line with related theories that postulate that schizophrenia may result from an impairment in the brain's predictive mechanisms.

A GENERATIVE MODEL FOR MULTISENSORY INTEGRATION

When we formalize brain function as hierarchical Bayesian inference, we assume that the brain learns the causal structure of the world. This causal structure is hierarchical (e.g., forest causing trees, tree causing leaves, etc.) and reflected in the cortical hierarchy from primary sensory areas to association areas. Inference corresponds to the inversion of this model, e.g., determining the most likely cause of the sensory evidence. In CI, this inference is implemented through belief propagation. For technical details about belief propagation with and without loops, please refer to the Supplementary Material and to relevant books and papers. Previous work on CI focused on simple generative models that consisted of a single stream (e.g., the pairwise graph: ForestTreeLeafColour green, see. Those simplified models can describe hierarchical processing within one single sensory modality but are insufficient to account for crossmodal phenomena. Here, we extend these generative models by considering two parallel hierarchies, each reflecting a different sensory modality (e.g., audition (birdsong) and vision (image of a bird)). The two modalities share a common node at the top, representing a common cause (singing bird). Through this node, stimuli from multiple modalities can be integrated (e.g., merging the auditory and visual inputs caused by the same bird). Moreover, one modality can affect the other (e.g., visual identification of the bird can also predict its song). The nodes within each of the two hierarchies can be interpreted as unimodal sensory areas representing features of increasing complexity, as in the ventral stream from V1 to V4 for the visual modality. The top node could correspond to higher-order association areas where multisensory integration occurs, such as the superior temporal sulcus or the occipital-temporal junction in the case of audio-visual perception. For illustration purposes, we will consider here the example of the stimuli triggered by a singing bird (see Figure). Inference in this graph works by iteratively calculating probabilistic messages and beliefs. In the absence of loops, sensory information climbs the cortical hierarchy, moving from sensory to association areas, and conversely prior information descends the hierarchy (in the opposite direction). In the current model, two parallel hierarchies can talk to each other via the top node. In other words, because of this (potential) binding, the presence of a stimulus in one modality increases the probability that there is also a stimulus in the other modality. Once the sensory information reaches the association cortex, it does not stop there but can enter the opposite hierarchy as a prior (Figure). In summary, each sensory modality normally receives three types of information, each coming from a different source: (i) its own sensory evidence, (ii) the sensory evidence from the other modality (computed as a prior), and (iii) prior knowledge that reaches the association cortex from the top., here we consider a model with two hierarchies, each representing a different sensory modality (e.g., audition (orange) and vision (green)).

THE TWO MODALITIES ARE CONNECTED THROUGH THE TOP NODE (GREY), WHICH CORRESPONDS TO THE LOCUS OF MULTISENSORY INTEGRATION (E.G., ASSOCIATION CORTEX). OUR EXAMPLE ILLUSTRATES HOW STIMULATION OF THE TWO DIFFERENT SENSORY MODALITIES MIGHT HAVE ARISEN FROM THE SAME "MULTI-MODAL" STIMULUS. INFERENCE CORRESPONDS TO THE INVERSION OF THIS FORWARD MODEL. (B):

A potential implementation of the generative model in (a) in the brain's hierarchical structure. According to Bayesian accounts of perception, the brain learns the causal structure of the world, which is represented in the cortical hierarchy. Filled nodes correspond to higher regions (the OFC, ACC and hippocampal complex), potentially sending different kinds of feedback information to the sensory association cortex.

SYNAESTHESIA, HALLUCINATIONS AND VISUAL ILLUSIONS

We simulated the CI framework to test the link between the meso-scale (i.e., probabilistic computations implemented by a message-passing algorithm) and the macroscale (i.e., phenomenological varieties of the psychotic experience under psychedelics and in schizophrenia). We used a graphical model composed of binary variables representing whether a feature is present or absent. Sensory evidence activates the nodes at the bottom of the hierarchy, while priors (e.g., expectation, memory, emotions, how common is this bird) can activate the top node. To study how the different modalities interact to produce unimodal or crossmodal aberrant experiences, we separately focused on belief formation within each modality. Two different models were implemented: i) A symmetrical model in which sensory hierarchies have the same structure and number of synaptic relays (Main Text); ii) An asymmetrical model in which anatomical differences (e.g., in the total number of levels or the number of synaptic relays between the receptors and the level in which the belief is constructed) generate modality-dependent aberrant experiences (e.g., mainly auditory hallucinations in schizophrenia; Supplementary Material; Figuresand) We tested the robustness of the results with different parameter values (i.e., weights (strength of feedforward and feedback connections), strength of the loops, likelihoods and priors). The results were obtained after 30 iterations of the algorithm, with each iteration corresponding to one exchange of messages in both directions between all the connected nodes. The simulations specifically covered three different scenarios: 1. In the "synaesthesia" scenario, we stimulated one modality (e.g., audition) with strong sensory evidence, while the other modality (e.g., vision) received weaker negative evidence. 2. In the "sensory-driven hallucination" scenario, both modalities were stimulated by noise, i.e., sensory evidence too weak to be considered reliable. In both cases, to avoid additional confounding effects, we did not consider any prior (𝐿𝐿 𝑃𝑃 = 0). 3. Finally, in the "perceptual illusion" scenario, the absence of sensory stimulation was contradicted by a strong prior for the presence of a bird.

DIFFERENT TYPES OF LOOPS FOR DIFFERENT CLINICAL PHENOTYPES? SCENARIO 1: SYNAESTHESIA

We first explored the symmetrical model and the scenario of synaesthesia (scenario 1). This is illustrated in Figure, with relevant beliefs shown in Figure, e (see also Figure). Here, the system receives a strong auditory activation from a birdsong (orange hierarchy) but no corresponding visual stimulation (evidence supports the absence of a bird; green hierarchy). When the system does exact inference (belief propagation without loops; Figure), those two pieces of sensory evidence climb their respective hierarchy, reach the association cortex (i.e., the grey node) and enter the opposite hierarchy where they are fed back as priors. As expected, such a system experiences the presence of a birdsong (𝑃𝑃(𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝑆𝑆 𝐴𝐴 , 𝑆𝑆 𝑉𝑉 ) ≫ 0.5), but the auditory feedback to the visual area is insufficient to overcome the evidence against the presence of a visual stimulus. Thus, the system does not "see" the bird, but only "hears" it (𝑃𝑃(𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏|𝑆𝑆 𝐴𝐴 , 𝑆𝑆 𝑉𝑉 ) < 0.5) (Figure, e; left bars for each modality). The addition of descending loops results in the reverberation of messages between the two modalities (Figure). Because of this cross-amplification, the modality receiving the strongest evidence can completely dominate the other. In visual areas, the auditory feedback overcomes the sensory evidence, and the system "sees" a bird even if its visual input says otherwise (Figure; right bar for each modality). In the auditory modality, we observe an over-confidence (the probability of the birdsong rises higher than it should, based on auditory information alone). This is as if the "imagined" visual bird came in support of the real (but not entirely reliable) auditory information. In short, the presence of descending loops enhances the communication between the different sensory modalities, which results in a concomitant experience in the second modality, a phenomenon that corresponds to synaesthesia. In contrast, the addition of ascending loops degrades the communication between sensory modalities (Figure). Information is amplified only within the modality of origin, while information from the two modalities cannot be integrated. This makes synaesthetic experiences almost impossible. In our example, the evidence for the birdsong is amplified within the auditory modality but not the visual modality, resulting in a strong unisensory percept (Figure, right bar for each modality).

SCENARIO 2: HALLUCINATIONS

In the second scenario, we explored hallucinations by considering whether strong beliefs could be generated in the absence of meaningful sensory information or prior. We thus tested the case when both modalities receive weak evidence fluctuating close to the chance level (e.g., as a consequence of sensory receptor noise, Figure). Without CI, all beliefs remain close to chance, as would be expected from a Bayesian integrator. In agreement with previous results, loops generate strong beliefs unrooted in the available sensory evidence. However, these beliefs have unique patterns for each type of impairment. Descending loops generate a strong, crossmodal and correlated experience (i.e., multisensory hallucinations; Figuresand). Both modalities strongly believe either in the presence of a singing bird (if, by chance, the total sensory evidence is slightly in favour of it) or in its absence (if the combined sensory evidence is slightly against it). When hallucinations occur, they are always audio-visual. In contrast, ascending loops result in segregated sensory modalities. Thus, if the input to one modality is by chance slightly in favour of a bird, it will develop a strong belief in that direction regardless of what is happening in the other modality. For example, the visual system can believe strongly that it sees a bird, but the auditory could believe equally strongly that it does not hear one. In that case, the two modalities develop inconsistent interpretations of the sensory input rather than an integrated global percept (i.e., unimodal hallucinations; Figuresand). Crossmodal hallucinations remain possible with ascending loops but only when unimodal aberrant beliefs co-occur by chance.

SCENARIO 3: ILLUSIONS

Finally, the third scenario probed the effect of a strong prior (Figuresand). If the sensory evidence contradicts the prior, this could correspond to perceptual illusions. If sensory evidence is absent, it could also account for mental imagery or what we could name a prior-driven hallucination. Importantly, descending loops amplify the prior, resulting in more illusions and stronger mental imagery (or prior-driven hallucinations; Figuresand). In contrast, ascending loops force the system to resort more to its sensory evidence, which leads to decreased susceptibility to illusions but less capacity to integrate information from multiple modalities (Figuresand). Overall, we show that while they can both lead to hallucinations, the two types of circularity generate very different phenomenological experiences. Descending loops enhance communication between sensory modalities (potentially between cognitive modules as well), leading to crossmodal hallucinations, synaesthesia, stronger mental imagery and visual illusions. On the other hand, ascending loops intensify segregation while also amplifying local sensory information, resulting in unimodal aberrant experiences, inconsistencies between beliefs in different pathways and less vulnerability to illusions. Interestingly, the former appears closer to the clinical properties of drug-induced psychosis (psychedelics, especially the DMT-containing brew ayahuasca), while the latter shares important properties with schizophrenia. In the next sections, we will specifically investigate the links between the meso-scale and the micro-scale (neural circuits), suggesting detailed implementations for the different types of loops. We stimulated one modality (e.g., audition) with strong, unambiguous information (e.g., birdsong; 𝐿𝐿 𝐴𝐴 = 2.9), while the other modality (e.g., vision) received negative evidence (absence of bird; 𝐿𝐿 𝑉𝑉 = -2). In the absence of loops, the system hears the birdsong (belief above 0.5) but is more uncertain regarding the presence of a bird (belief below but close to 0.5). Because of the cross-amplification caused by descending loops (d), the prevailing auditory information pushes both beliefs towards 1, eliciting an inversion in the case of vision. Thus, the system hears a birdsong and perceives the image of a bird, despite only audition being stimulated (synaesthesia). In contrast, ascending loops cannot generate such an inversion (or synaesthesia), because self-amplification inside the visual hierarchy reduces the visual belief towards 0 (e). In the presented simulations, both weights (prior and sensory weights; they correspond to conditional probabilities that quantify the strength of the reciprocal causal links between connected nodes) were taken equal to 0.95.): Results of simulations for illusions / mental imagery / prior-driven hallucinations (scenario 3). In scenarios 1 and 2, the two hierarchies are stimulated only by sensory inputs (prior knowledge is absent: 𝐿𝐿 𝑃𝑃 = 0). In scenario 3, we investigate the effect of such a prior. In this particular example, we consider a prior that contradicts the sensory stimulation (𝐿𝐿 𝐴𝐴 = 𝐿𝐿 𝑉𝑉 = -1.4; 𝐿𝐿 𝑃𝑃 = 1). Because the (contradicting) prior is weaker than the sensory information, in the absence of loops, both beliefs are below 0.5 (both the bird and the birdsong are absent; note that beliefs are equal in the two modalities because the two hierarchies are identical

SYMMETRICAL MODEL] AND THEY RECEIVE EQUALLY STRONG STIMULATION). (C) WHEN THE PRIOR IS AMPLIFIED (DESCENDING LOOPS), INFERENCE IS DOMINATED BY THE FEEDBACK, RESULTING IN BELIEFS CLOSE TO 1 (THAT COULD CORRESPOND TO AN ILLUSION, ENHANCED MENTAL IMAGERY OR A PRIOR-DRIVEN HALLUCINATION, DEPENDING ON THE CONTEXT). (D)

Ascending loops (amplification of sensory information) have the opposite effect (less vulnerability to illusions, weaker mental imagery and no prior-driven hallucinations). In all the presented simulations, both weights (prior and sensory weights) were taken equal to 0.95.

LOOPS ARE PREVENTED BY DIFFERENT TYPES OF INHIBITION

We have highlighted that hierarchical inference depends entirely on a precise cancellation of the reverberated information, which prevents the formation of loops. Previous formulations of the CI algorithm considered a correction at the level of the messages (eq. S2). Here, we suggest a novel formulation of the algorithm in which the correction takes place at the level of the beliefs (see Supplementary Material for further details). For a pairwise graph, the resulting beliefs can be written as follows: Messages are simply sigmoid functions of the beliefs of the sending nodes (without correction; see Supplementary Material), while the last two terms correspond to the subtraction of the redundant sensory (𝑓𝑓 term) and prior (𝑏𝑏 term) information. The neural interpretation of eq. 1 is straightforward. Belief at level 𝑏𝑏 is generated by integrating excitatory inputs from the levels above and below, which are balanced by inhibitory inputs from interneurons at the same level. This inhibition is driven by excitatory neurons at the same level and the levels above or below (Figure). Crucially, the prior term (f) depends on beliefs at the same level and the level below. Conversely, the sensory term (g) depends on beliefs at the same level and the level above. This distinction speaks to an important anatomical difference between the two inhibitory mechanisms: interneurons removing descending loops are driven by lateral and feedforward connections; conversely, those responsible for control of ascending loops are driven by lateral and feedback connections (Figure).

A CANONICAL MICROCIRCUIT IMPLEMENTING CIRCULAR INFERENCE IN THE SENSORY CORTEX

The cortex is widely viewed as a hierarchical structure, whose networks are organized in a laminar-specific manner, leading to the notion of "canonical" microcircuits. Those repeating circuits have long been viewed as the basis of many cortical computations. What is the structure of those microcircuits and how are they linked to the CI framework? In a cortical microcircuit, both pyramidal cells and interneurons play important roles and exhibit strong laminar specificity. Long-range connections consist of pyramidal cell axons, mostly targeting other pyramidal cells. According to the dominant view, feedforward information originates from L2/3 pyramidal cells (and thalamus) and mainly targets L4, with projections onto both pyramidal cells and interneurons in a nonselective manner. These neurons then project to superficial layers; but see also, and from there, information reaches deep layers, especially L5. Reverse inter-laminar connectivity within an area (e.g., dashed line from L5/6 to L4 in Figure) is less frequent. Nevertheless, strong connections exist between L5/6 pyramidal cells and L4 interneurons. In the opposite direction, feedback is less laminar specific: feedback information originates predominantly from the deep layers and targets all layers, except L4but also non-specific thalamic nuclei. Importantly, many feedback connections terminate on interneurons in L2/3but also in L1, which then form reciprocal connections with pyramidal cells in superficial layers. Figureillustrates a (simplified) canonical microcircuit implementing CI. We suggest that pyramidal cells in the superficial layers act as integrators, receiving all the available information and generating the beliefs. This description of the cortical microcircuits illustrates in a dramatic way how the recurrent connectivity of the brain can generate information loops if inhibition fails to balance excitation (Figure). In particular, feedback information re-climbs the hierarchy, generating a descending loop. For example, information sent from the superficial layers of V2 to the granular layer of V4 will be sent back due to the internal structure of the cortical column. Similarly, sensory information forms a positive feedback, involving cortical or thalamo-cortical ascending loops. More importantly, this illustration gives crucial hints about the implementation of the inhibitory mechanisms controlling the propagation of information. As described before, descending loops are balanced by inhibition driven by feedforward excitatory inputs. This description fits nicely with L4 (and potentially deep layer) interneurons (Figure). Hypo-activation of those interneurons (e.g., because of aberrant modulation of deep layers by serotonin via 5HT 2A , as observed with psychedelics) would lead to dis-inhibition of this part of the cortical circuits, resulting in an amplification of top-down messages. Likewise, ascending loops could be balanced by feedback-driven inhibition. This description points to L1 interneurons, with the possible involvement of L2/3 interneurons as well (Figure). Impairments of inhibition in superficial layers (e.g., due to dopaminergic abnormalities in schizophrenia) would cause amplification of sensory information and thus more segregation of the sensory modalities. Note that this suggestion is compatible with the influential "dysconnectivity hypothesis" of schizophreniaand especially with a variation of this theory implicating thalamo-cortical loops.

WHICH CAN BE AVOIDED IF INHIBITION SUCCESSFULLY REMOVES ALL REDUNDANT INFORMATION (BALANCES EXCITATION). INHIBITION

driven by bottom-up information, mediated by interneurons in L4 (potentially also interneurons in L5/6), removes descending loops, whereas feedback-driven inhibition, mediated by interneurons in L2/3 (and/or L1), is responsible for ascending loops.

GENERAL DISCUSSION

The goal of this paper was to delineate a CI-inspired unifying and multiscale theoretical account of false perceptions, including unimodal hallucinations (e.g. auditory-verbal hallucinations in schizophrenia), multimodal hallucinations / synaesthesia (exemplified by serotonergic agonists such as DMT) and illusions. Overall, we argue for a link between the macro-scale (behavioural and phenomenological experience), the meso-scale (belief propagation and CI in neural networks) and the micro-scale (E/I networks in a canonical cortical microcircuit. Previous work has linked ascending loops with psychotic symptoms, including auditory hallucinations, persecutory delusions, and jumping-to-conclusions bias, and low vulnerability to illusions in schizophrenia. Additionally, we have suggested that mild (descending) loops might play an important role in normal brain functionand underlie common perceptual phenomena such as bistable perception. Here, we extended those ideas by showing that different loops can generate very different aberrant perceptual phenomena. We notably showed that descending loops lead to over-integrated sensory hierarchies that result in crossmodal hallucinations, synaesthesia, illusions and increased mental imagery, all common features in psychedelic-induced psychosis (other common properties such as the effect of set or the effect of emotions on perception, could also be explained by amplified top-down effects). We conclude that while ascending loops might be a prominent impairment at the roots of schizophrenia symptoms, descending loops could underlie the rich phenomenology induced by serotonergic agonists such as. Interestingly, this unifying framework is related to a number of different theories that have addressed the problem of psychosis, many of which were built on the idea that hallucinatory phenomena result from impairments in predictive mechanisms of the brain. In one study, the authors suggested that enhanced priors, mediated by over-activation of deep layers, generated the subjective effects associated with psychedelics, while in another study, the same effects were associated with impaired bottom-up processing (combined with intact top-down processing) mediated by enhanced AMPA signalling. The present account also appears compatible with another contemporary theory of psychedelics, the entropic brain theory (EBT). EBT suggests that psychedelics increase the entropy of brain activity, rendering it more chaotic and susceptible to intrinsic and extrinsic influences, while psychedelics also increase connectivity between resting-state (and task-positive) brain networks, in agreement with the enhanced integration induced by descending loops. Note, however, that the CI framework, contrary to EBT, is a functional theory directly derived from normative principles. Our findings have important implications for our understanding of the neural mechanisms that implement and control perceptual inference. We delineated a canonical microcircuit implementing hierarchical inference (see also, for a related microcircuit implementing predictive coding). Inhibitory connections are crucial in preventing CI or in let CI happen, depending on how these inhibitory connections are modulated. We argued that feedback-driven inhibitory interneurons situated in superficial layers (L2/3 and/or L1) mediate the control of the ascending loops. Conversely, inhibition in the deeper layers (L4 and/or L5/6), driven by feedforward information, is mainly responsible for regulating descending loops. Having established theoretical ties between psychedelics and descending loops, as well as between schizophrenia and ascending loops, it's tempting, and perhaps not without some merit, to speculate about the role of neuromodulation in CI. The involvement of serotonin (through the action of the 5HT 2A receptors) in the psychotic effects of psychedelicssuggests a link between serotonin and the part of the circuit controlling descending loops. Similarly, dopaminergic involvement in schizophreniasuggests that this neuromodulator (in tandem with acetylcholine) could target the part of the circuit controlling ascending loops. Importantly, the CI framework makes a number of new testable predictions. First, the CI framework offers a tentative explanation for the clinical and neurobiological variability observed in psychosis. Indeed, although a majority of patients with schizophrenia experience auditory hallucinations, approximately 30% of them also experience both auditory and visual hallucinations. Additionally, even if most of the patients respond well to typical antipsychotic medication (dopamine (DA) antagonists), one in four exhibits refractory hallucinations. Crucially, most of these "treatment-resistant" hallucinations still respond well to clozapine, an atypical antipsychotic characterized by a high affinity for serotonin receptors. Although evidence for a link between these two groups of patients (i.e., those exhibiting complex multisensory hallucinations and those exhibiting drug-resistant hallucinations) is currently very sparse, it is tempting to suggest that these differences could be due to different types of CI. Most schizophrenia patients, who exhibit dopaminergic (and perhaps cholinergic) dysregulation, could be predominantly impaired in their ascending loops (resulting in an overinterpretation of sensory evidence and generally weakened top-down effects). This form of schizophrenia (type A, according to) could be modelled by the psychotomimetic effects of amphetamines (DA agonists) and improved by first-line antipsychotic drugs (mostly DA D2-receptor antagonists). In contrast, a minority of patients with associated serotonergic impairments could be affected in their descending loops (resulting in strengthened top-down effects and excessive crossmodal integration). This would lead to more audio-visual and drug-resistant hallucinations, but better responses to agents with serotonergic properties, such as clozapine or even ketanserin and pimavanserin. This category of patients could constitute a good candidate for type B schizophrenia. Some recent case studies support this claim. The aforementioned prediction also calls for an important methodological comment. Since psychosis is not a monolithic experience, a better clinical characterization is crucial, especially when building computational models with clinical applications (computational assays;. If not taken into account, this variability could, at minimum, contaminate the results and lead to contradicting evidence (e.g., more or less susceptibility to illusions) and, at worst, lead to inefficient treatments. In particular, despite the importance of simultaneous multisensory hallucinations as a potential diagnostic tool, only a few studies have systematically studied them. As a result, it is difficult to objectively evaluate their prevalence (as opposed to serial hallucinations), both in schizophrenia and under psychedelics. A second important prediction of the model comes from the fact that descending loops cause over-integration and thus amplification of information in both modalities. As presented in Figure, this results in a general over-confidence, affecting both modalities. Interestingly, this is a unique prediction since different models (e.g., those based on increased prior weights; not presented here) would only generate over-confidence in the nonstimulated modality and an under-confidence in the stimulated one; future studies, based on carefully designed behavioural tasks involving stimulation of more than one sensory modalities, will have to scrutinize and arbitrate between those competing predictions. Finally, another testable prediction pertains to the suggested implementation of CI, in particular the laminar and input specificity of inhibition and its connection with schizophrenia (unimodal hallucinations) and drug-induced psychosis (multimodal hallucinations and synaesthesia). Although standard imaging techniques do not possess the necessary spatial resolution to test so precise predictions, recent advances in high-field laminar fMRI could make it possible to probe implementations of message-passing algorithms at the level of microcircuits. We need to acknowledge some limitations. First, a cautious approach is needed regarding the potential neural substrates of CI. For instance, the suggested microcircuit is necessarily simplified, ignoring less common connections, interneuron specificities (e.g., differences between fast-spiking interneurons and adaptive interneurons) and within-layer details (e.g., detailed connectivity within L2/3).. Second, the current functional account of synaesthesia constitutes a model of druginduced synaesthesia and it was not designed to account for developmental synaesthesia (e.g., grapheme-colour synaesthesia, experienced by a small number of people without any drug consumption). That said, it is difficult to disregard the similarity between the descending loops (amplification of priors) and ideas such as the "disinhibited prior" proposed for developmental forms of synaesthesia. Overall, we put forward a unifying, transnosographic and multiscale account of psychosis, with a special focus on psychedelics. This approach may pave the way for further investigations, such as accounting for different psychotic experiences and/or contexts of occurrence (e.g., hallucinations in non-clinical populations or people with Parkinson's disease), examining the potential neurophysiological signatures of CI in humans and linking the results to animal models. First, the visual modality (in green) contains more nodes (more synaptic relays between the sensorium and the sensory association cortex) compared to the auditory modality (in orange). Second, visual beliefs are generated further from the sensorium (and closer to the association cortex). Taken together, these differences result in stronger descending loops in the visual hallucinations can be generated with the asymmetrical model due to the weak amplification of visual information within the visual modality (𝐵𝐵 𝑉𝑉 is generated at a high level, close to the sensory association cortex). In the presented simulations, both weights (prior and sensory weights) were taken equal to 0.95. Interestingly, the asymmetric model predicts asymmetric effects of the loops: auditory inputs prevail over visual inputs in the case of descending loops (b) and auditory but not visual amplification in the case of ascending loops (c). Both weights (prior and sensory weights) were taken equal to 0.95.