Minds and Brains, Sleep and Psychiatry

The neuronal doctrine of Ramon y Cajal (27) and the reflex doctrine of Charles Sherrington (28) dominated neuroscience from 1890 until about 1950. In the absence of sleep and dream science, Sigmund Freud was unable to complete his Project for a Scientific Psychology in 1895. Instead, he turned his attention to the psychanalytic interpretation of dreams (29), which he insisted was in no way neurological (30). The resulting split in mind‐brain unity had profound effects on medicine, philosophy and psychology. One might contend that this split perpetuated the dualism of Descartes and that Cartesian Dualism continues to divide concepts and fields which ultimately belong together (31).

While psychiatry and neurology proceeded along parallel but separate tracks from the 1890's, neuroscience flourished with the elaboration of principles, such as synaptic action, neurotransmission, and the chemical mediation of neuronal signaling. Only in about 1950 did it became possible to record from individual neurons in living animals (32). A conceptual framework for experimental brain–mind integration was concurrently provided by the discoveries of brain activation in waking (33) and in REM sleep dreaming (34). These discoveries resulted from the technological advance of recording the electrical activity of the brain via the electroencephalogram (EEG) (35).

These neuroscientific advances led to some important insights: The localization of REM sleep control to the pontine brain stem (36), the modulation of cerebral activation by aminergic brain stem neurons (37), the mediation of REM sleep events—including REMs themselves—and the inhibition of motor tonus by specific neurophysiological mechanisms (38). Extracellular neuronal recording produced neurophysiological data from REM sleep that was subsequently integrated with quantitative analysis of dream mentation, to create the first brain‐based theory of dreaming (20, 30). Over the subsequent four decades these findings have been confirmed and enriched (39, 40).

The psychologist Ursula Voss used EEG to show that when human subjects became aware that they were dreaming—instead of erroneously supposing themselves to be awake as is usual in REM sleep—they exhibited the frontal lobe activation, normally seen in waking—together with parietal and occipital signs of REM sleep (41): see Figure 1. These findings were later supported by a case study using combined EEG and functional magnetic resonance imaging (42), showing heightened BOLD activation in frontal areas, in addition to the precuneus and inferior parietal lobules; all of which are implicit in self‐referential processing and the experience of agency (43).

This finding was supported by a subsequent study showing heightened activity in the dorsolateral prefrontal cortex during dream lucidity (44) and a further study showing elevated resting‐state functional connectivity between the prefrontal and temporoparietal association areas in frequent lucid dreamers (45). Lucid dreaming thus reveals itself to be a hybrid state of the brain–mind, with both waking and dreamlike features—with its phenomenal aspects being reducible to differential activity in specific brain structures (46). This suggests the brain and the mind operate as one, with even the highest features of conscious activity—such as self‐referential processing and insight—having definitive and functionally segregated neuronal correlates.

A related point is that the brain–mind is not necessarily unified with respect to its own cardinal states. The expression “I am of two minds” reflects the ambivalence (or multivalence) of cognition (and emotion) in waking consciousness. The locality of brain activation has been clearly demonstrated in studies of regional cortical activation in NREM sleep (47) and is often referred to as “local” sleep (48, 49, 50). Normal and abnormal conscious experience may thus be not only linked in an intimate way to neuronal dynamics but show the same kind of segregation and differentiation seen in the neurophysiological correlates of sleep (47, 51, 52).

The Helmholtzian pillar of the current thesis associates unconscious inference with perception and sentience (4, 6). Formal treatments of this notion rest upon predictive processing; namely, using sensory impressions to confirm or disconfirm hypotheses about how our sensory inputs are generated (53). When sensations are generated actively, the accompanying predictions must reflect the way that the sensorium is sampled. Until recently, the electrical encoding of internal predictions that accompany eye movements had only been demonstrated in cats (11, 36). The psychiatrist Charles Hong used brain imaging to demonstrate robust evidence for the existence of these internal stimuli in human REM sleep (54). Each and every human eye movement command—from the pontine brain stem to the forebrain—is associated with information transfer that assures perceptual integrity in waking. And which may be used to construct the visual imagery of dreams. Hong et al (55) have recently discussed these possibilities and how they might be pursued to establish the unity of the brain–mind in visual perception, be it external (as in waking) or internal (as in dreaming). Much of this empirical work was motivated by the AIM model of sleep which we now briefly summarize (see also Figure 2).

The activation‐synthesis hypothesis of dreaming (20, 21, 30) and the reciprocal interaction model of sleep cycle control (20) have been modified, simplified and extended to accommodate altered states of consciousness: see Figure 2 and (13). The ensuing AIM account tries to integrate neurophysiology and psychology as follows. The three dimensions of AIM state‐space represent activation (A), input source (I) and modulation (M), as measured neurophysiologically. The discovery of REM‐locked pontine activity in humans (54) enriched the AIM model by providing evidence for human ponto‐geniculo‐occipital (PGO) wave activity (56) that had only been definitively recorded in the pons (57). Figure 3 provides a schematic based on the AIM model in Figure 2 that focuses on the functional anatomy of sleep in terms of neuromodulation. This account is scaffolded on functional brain states and neuromodulation at the level of synaptic physiology, which brings us to the psychopharmacology of sleep—and its intersection with psychopathology.

The aminergic and cholinergic neurons that underwrite conscious states are affected by drugs of common use in medicine. These include atropine and the antidepressants (especially the amine reuptake blockers). Because such drugs have both short and long term effects they are useful as sedatives and mood regulators, depending on dose; for example, (58). This is not surprising given the well‐known intimacy of sleep and affect, where the most prominent psychiatric beneficiary of sleep science is mood disorder (59). The reduction in REM sleep latency seen in depression predicts positive response to amine reuptake treatment (60, 61). This speaks to the role of classical ascending modulatory neurotransmitters in the control of both sleep and mood. In the next section, we take some of the above fundaments of sleep and dream science and see how they relate to formulations of the sentient brain, in terms of predictive processing and belief updating. We will see that neuromodulation is a recurring theme that may have a special relevance for psychiatry.

There are several ways in which we can license a focus on precision and the encoding of uncertainty in active inference. We will briefly rehearse a few prescient approaches, to show that they all converge on the same conclusion; namely, inference—and especially false inference—depends sensitively on getting precision right (78). Getting precision right depends upon the right deployment of neuromodulatory mechanisms that coordinate message passing in cortical and subcortical hierarchies (76, 79).

First, from a purely technical perspective, a commitment to a dual aspect monism requires the existence of a dual aspect information geometry that links neuronal dynamics to belief updating (10). The technical details are beyond the scope of this essay; however, they can be understood heuristically in terms of two geometries that describe the statistics of neuronal activity (and connectivity) on the one hand, and statistics of (Bayesian) beliefs encoded by that activity (and connectivity) on the other. In other words, there are two information geometries that inherit from the ensemble dynamics of neuronal populations. The first pertains to the probabilistic evolution of neuronal states, while the second pertains to the evolution (i.e., updating) of beliefs about something—beliefs that are encoded by neuronal states (and connectivity). Crucially, these two information geometries arise in any system that self‐organizes itself in a way that can be separated from its environment (80, 81, 82).

This may sound rather abstract; however, it is just a statement of a fact—for which we are our own existence proof. Namely, the neurophysiology and neurochemistry of our brains conforms to the laws of thermodynamics and an associated information geometry (83). At the same time, these physiological states (c.f., the AIM model) are necessarily equipped with an information geometry that pertains to things and narratives “out there.” Dreaming is a beautiful example of this: my brain enters a particular sleep (e.g., REM) state, characterized by a particular physiology, while at the same time I dream about things in my lived world. Crucially, the things I dream about are essentially grounded in my navigation of that world. In other words, REM dreams are always animated (hence in the service of running motor programs offline), emotional (with anxiety, aggression, and elation predominating), bizarre (meaning only remotely associative) and so on (13, 84). “Chase” dreams are common and well‐known, and frequently overlap with—and merge—into nightmares. In the dream state, one rehearses threatening situations (running away, turning, and fighting) but also simulates more abstract potentialities. One can fly or make love (85), particularly when lucid (86).

Technically, information geometries underwrite a (Bayesian) mechanics of the self‐organizing brain and thereby characterize the nature of belief spaces (82). This can be appreciated from a number of complementary perspectives. For example, one can apply gauge theoretic treatments (87). Alternatively, one can think about the notion of distance and information length—as measures of how far beliefs move during inference or belief updating (83, 88). All these treatments converge on the same quantity; namely, the precision of beliefs—that is, the confidence or inverse variance of a belief distribution.

Information geometry allows us to conceive of belief‐spaces in which two beliefs are equipped with a measure of distance between them. The key notion here is some state‐space (e.g., the physiological states of the AIM model), where every point in this space corresponds to a Bayesian belief (i.e., a probability distribution). This special sort of state‐space is called a statistical manifold. For example, imagine I saw something fluttering out of the corner of my eye. On foveating the source of “fluttering,” I “see” it is a butterfly. This active sampling of the visual scene moves my (prior) beliefs that there was a small creature out there—with probability mass distributed over all kinds of plausible alternatives (e.g., insect, bird, leaf, etc.)—to a precise (posterior) belief (e.g., butterfly). This belief updating corresponds to moving from one point on a statistical manifold to another. So, what are the key determinants of that movement and how is movement on a statistical manifold measured?

In statistics, this measure is based on the Fisher information metric. Under mild (i.e., Gaussian) assumptions, this metric corresponds to the precision or confidence associated with (probabilistic or posterior) beliefs. Intuitively, we can envisage a myriad of beliefs occupying some belief space, all attracting each other to a greater or lesser degree. This is very much like a celestial n‐body problem, in which heavenly bodies, pull each other in different directions—to find their minimum free energy configuration. In this intuitive physics, precision corresponds to the mass of each belief (or associated prediction error). In other words, a belief (or associated prediction error) that is afforded greater precision will behave like a massive body—attracting smaller beliefs to it—and becoming relatively impervious to others. For example, if sensory precision were inordinately high, believes based upon sensory evidence would acquire a gravity that pull other (empirical prior) beliefs towards it. This attraction is the manifestation of (mindful) forces that underwrite belief updating from prior to posterior Bayesian beliefs, that is, before and after the effects of sensory forces. Conversely, if prior precision is high the forces exerted by sensory impressions exert less influence on beliefs deep within a hierarchical belief system. These two extremes are just the difference between waking and sleep, respectively; in waking we have information from the outside world to shape perception, whereas in sleep, we are sequestered from these sensory impressions.

On this view, one can see how getting precision right becomes crucial: precision is the glue, or bridge that realizes conditional dependencies among distinct beliefs about the lived world. If this coupling were to fail, there would literally be a disintegration of the psyche and loss of central (deep) coherence that characterizes conditions like schizophrenia and autism (89, 90, 91, 92). Indeed, as we will see below, any form of belief updating must have precision at its heart, suggesting that psychopathology is a pathology of precision.

From the perspective of predictive coding (a canonical scheme for predictive processing) (93, 94), precision is the key quantity that coordinates belief updating by differentially weighting prediction errors as they ascend cortical hierarchies (75, 95, 96, 97, 98, 99). In these schemes, prediction errors represent the difference between sensory input and top‐down predictions of that input. Any mismatch then revises Bayesian beliefs or expectations encoded by neuronal activity to implement belief updating—and provide better predictions that resolve prediction errors (see Figure 3). In engineering, precision is known as Kalman gain (100) and scores the precision or reliability of prediction errors; such that only precise prediction errors have access to higher levels of belief updating (101). In this setting, it can be seen that getting the precision or gain right is crucial for veridical inference.

From a physiological perspective, precision is thought to be encoded in the synaptic gain or excitability of particular neuronal populations (e.g., superficial pyramidal cells encoding prediction errors). It can be subsumed under notions of excitation‐inhibition balance or cortical gain control at the population level (102, 103). At the synaptic level, it encompasses a broad range of neuromodulatory mechanisms that generally involve fast‐spiking inhibitory interneurons that Interact with pyramidal cells (104, 105, 106, 107, 108). In turn, these (often synchronous) interactions depend upon N‐methyl‐D‐aspartate (NMDA) receptor function and, crucially, classical ascending modulatory neurotransmitter systems (e.g., noradrenaline, serotonin, and acetylcholine) (109, 110).

From a psychological perspective, precision manifests as various forms of attentional gain control or selection (101, 111). The right deployment of sensory precision in the visual domain can be thought of as mediating spatial attention, whereas higher in the hierarchy it can be associated with feature attention and, possibly, internal attention states (112, 113).

It is at this point that we see the formal relationship between precision and the AIM model (21). The activation (A) of certain neuronal populations—that entail belief updating—depends upon their input (I), which is under modulatory control (M). The permissive requirements for sleep rest upon a selective disconnection of sensory inputs via aminergic neuromodulation, which can be thought of as a physiologically mediated inattention to the sensorium. In many respects, it is the exact complement of mindfulness, in which sustained attention to a specific sensory input is maintained (114).

Finally, from a more philosophical perspective, the hierarchical control or predictions of precision underwrite notions of mental action and the distinction between the phenomenal transparency and capacity (68, 115, 116, 117). When associated with interoceptive inference, high level representations—that themselves predict the precision of lower‐level representations—may have a crucial role in emotional inference and elaborating a minimal selfhood (98, 115, 118).

In summary, a ubiquitous and key aspect of belief updating, or active inference, is the mechanics of encoding uncertainty, by optimizing (Bayesian) beliefs about precision. Mechanistically, this speaks to a key role of neuromodulation and the context‐sensitive control of synaptic efficacy. Psychologically, it ties in attention to any consideration of mindful processing (i.e., inference). Finally, these mechanisms are evinced most powerfully by the greatest contrast of sentience we know; namely, the difference between waking perception and dreaming. We now turn to aberrant precision in neuropsychiatry.

In the past years, nearly every psychiatric condition has been considered under the lens of precision; ranging from autism through to schizophrenia; from tremor through to Parkinson's disease; from obsessive‐compulsive disorder through to post‐traumatic stress disorder; from depression through to chronic stress; from fatigue through to functional medical symptoms (65, 97, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130). Table 1 provides a selective survey of some key publications, all pivoting on aberrant precision control. Each entry in Table 1 deserves its own discussion, showing how the same fundamental deficit in aberrant precision or attentional processing can manifest in domains as diverse as motor control to depersonalization. In what follows, we will focus on a couple of cardinal examples.

The thesis developed here helps clarify why the AIM model translates so gracefully when describing psychopathology and other disorders of movement and perception. This follows from the simple observation that if psychology (i.e., the mind) and physiology (i.e., the brain) inherit from the same (Bayesian) mechanics, it follows that psychopathology must be a pathology of inference. The particular kind of false inference—implied by the key role of precision—is then grounded in synaptic gain control that selects the sensory inputs (and, in predictive coding, ascending prediction errors throughout cortical hierarchies) responsible for inference or belief updating. This immediately suggests that any pathophysiology that can be framed as a synaptopathy, which interferes with synaptic gain control, can be understood in terms of aberrant precision control (131). We will focus on a couple of canonical examples to illustrate the crosscutting themes.

Perhaps the poster child for this line of thinking is autism, in which the belief updating has been proposed to rest on an imbalance between sensory evidence and prior beliefs—mediated by their relative precision (91, 132, 133, 134). The synaptic mechanisms that underwrite this imbalance are much less well developed but a compelling story can still be told at the level of psychopathology. Schizophrenia, has, arguably, the more advanced story—of abnormal synaptic gain control—to accompany the concomitant failures of belief updating and inference (64, 125, 135, 136, 137, 138, 139, 140). As with many psychiatric syndromes, the story starts with a failure of sensory attenuation—in the sense of a failure to modulate sensory precision (141, 142, 143, 144, 145, 146, 147). This explains some key soft neurological signs in schizophrenia such as an attenuation of the mismatch negativity, failures of slow pursuit eye movements and failures of sensory attenuation that underlie the force matching illusion (138).

In many instances, these examples demonstrate a paradoxical veracity of perception that renders it difficult to elicit illusions in schizophrenia (137, 148). The evidence for this kind of (paradoxical) deficit usually rests upon a careful use of psychophysics and computational modeling to illustrate aberrant precision control; for example (149). The physiological concomitants are easily articulated in terms of neuronal processing (e.g., abnormal excitation‐inhibition balance) (150), right down to the receptors that may mediate this control (e.g., NMDA receptors and GABAergic neurotransmission) (103, 110, 136, 151, 152, 153, 154, 155, 156, 157, 158). In this sense, schizophrenia may be a paradigm example of a (synaptic) disconnection syndrome that is largely manifest in terms of abnormal perception and belief updating (i.e., hallucinations and delusions). If this is the right story, then one would expect to see very similar phenomenology in other conditions that compromise synaptic gain control. Key examples here are synaptopathies; for example, Lewy body disease leading to hallucinosis (25, 159).

One key question in schizophrenia research, in this setting, is the emergence of delusions—which at first glance seem a prime candidate for high‐level belief systems that are afforded “too much” precision. This appears to fly in the face of a failure of sensory attenuation (i.e., too much sensory precision). Although consensus has not yet emerged, this might be an interesting example of a (perceptual) paradoxical lesion (160, 161, 162). In other words, a synaptopathy at higher levels of cortical (hierarchical) processing that partially reverses the effects of a primary insult at the sensory level. In other words, a compensatory increase in precision of higher level belief systems may be necessary to balance a loss of attenuation at lower levels; thereby engendering delusional belief systems as the best Bayesian explanation for sensory evidence—that one has lost the capacity to ignore. This clearly has close relationships with earlier notions, such as the aberrant salience hypothesis (163, 164) and fits nicely with the effects of psychedelic and psychomimetic drugs (165).

From a practical point of view, the foregoing suggests that pharmacotherapy—targeting neuromodulatory systems—is exactly the right way to proceed. Having said this, the multiple factors involved in synaptic gain control make psychopharmacology a challenging avenue to remediate a loss of delicate balance in hierarchical neuronal message passing. On a more positive note, if one subscribes to the above story, then the top‐down control of precision weighting becomes another therapeutic target. In this sense, it speaks to the possibility of using psychotherapy in conjunction with psychopharmacology (166, 167, 168), to improve a patient's ability to orchestrate the many facets of attention.

In concluding this section, we briefly consider movement disorders. The link between movement disorders and false inference is mandated by active inference accounts of how we sample our world. There are many examples here; for example, the difficulties schizophrenic patients have with slow pursuit eye movements (125, 169). Perhaps a more fanciful example would be Parkinson's disease. However, the inferential (mindful) mechanics could be exactly the same as seen in other conditions ranging from alpha synucleinopathies (76, 158, 170) through to autonomic dysfunction (123, 171).

The idea here is that our motor and autonomic reflexes are just in the service of fulfilling top‐down predictions (97). This tenet of active inference inherits from ideomotor theories and 20th‐century formulations such as equilibrium point hypothesis (172) and perceptual control theory and the passive movement paradigm (173, 174). In brief, if all our (motor and autonomic) actions are realizations of descending (proprioceptive and interoceptive) predictions, then the precision afforded descending predictions, relative to the peripheral prediction errors that drive reflexes, becomes crucial. Indeed, Parkinson's disease could be regarded as a complete failure to attenuate proprioceptive precision, such that the intention to move is immediately subverted on the basis of (unattenuated) proprioceptive evidence one is not moving (138). The neuromodulatory correlates—on an active inference reading of motor control—of this putative failure are well characterized in terms of dopamine neurotransmission (175)—and, indeed, the electrophysiological correlates of precision control such as beta activity (176).

In summary, a broad range of psychiatric and neurological disorders may yield to a coarse‐grained but mechanistic explanation in terms of aberrant precision control. Crucially, the mechanics at hand are of two sorts. On the one hand, there is the Bayesian mechanics of belief updating, which underwrites our perceptions and experience of the world—and how it is actively sampled—on the other hand, the synaptic mechanisms are, at a physiological, computational and population level, clearly situated in terms of their role in belief updating. In short, the encoding of uncertainty or precision in the brain inherits from a dual aspect monism.

AIM is a preliminary model, which should be regarded as a demonstration project. Its limitations with respect to the many already well‐known alterations of conscious state have been emphasized previously. One must also appreciate the anatomical problems associated with regional brain differentiation. To be true to reality, AIM—or its derivatives—should describe brain–mind conditions at all scales, not just at the level of the whole system. Is consciousness really a winner‐take‐all phenomenon or does it consist of multiple substrates, each with its own AIM? And speaking of AIM, are three dimensions adequate to define state‐space? Almost certainly not. Skeptics ask for many more and nay‐sayers believe the brain–mind problem to be beyond the reach of science. Others may consider themselves to be descendants of Lucretius, Aristotle—or even Spinoza—and there is a clear invitation to join in this glorious conceit.