2.1 Predictive processing

On the predictive processing view, the brain has stored prior beliefs (in the form of probability distributions) about the causes of sensory inputs in the world (Clark, 2013; Wiese & Metzinger, 2017). Prior beliefs are hierarchically organised, where higher-levels encode predictions about representations at lower levels. Prediction errors, arising from the discrepancy between the low-level predictions and incoming sensory signals, are passed up the hierarchy, where higher-level predictions are updated to minimise further prediction errors. Perception, then, both exteroceptive and interoceptive, is the product of (approximate) Bayesian inference, whereby the influence of prior beliefs and sensory evidence are weighted according to “expected precision”, e.g. confidence in the given context, to generate a posterior. Inference in these schemes is thought to occur across a hierarchy of inferred causes, where higher levels encode regularities that occur at larger spatial and temporal scales (Kiebel, Daunizeau, & Friston, 2008). In perceptual inference, sensory prediction errors can be minimised by tweaking the parameters of the generative model – that is, generating predictions to quash the influx of prediction error. Prediction error can also be minimised though action by changing the incoming sensory data to fit a prediction – for instance, I can move my eyes to bring my coffee cup into view, to fulfil the prediction of a coffee cup. Actions can be thought of as the fulfilment of proprioceptive (or oculomotor) predictions – an intended movement occurs as a result of predicting the proprioceptive consequences (Friston et al., 2010; Shadmehr, Smith, & Krakauer, 2010). There are detailed accounts of the neural implementation of these schemes available (Bastos et al., 2012; Keller & Mrsic-Flogel, 2018; Shipp, 2016).

2.2 Precision-weighting

A key feature of predictive processing schemes is the contextual flexibility afforded by precision-weighting (Clark, 2013; Feldman & Friston, 2010). Precision regulates the interaction between top-down and bottom-up signals, through the synaptic gain on neuronal populations signalling prediction error, in order to approximate optimal inference over time. Precision can be thought of as tracking both the reliability and relevance of the incoming sensory information, where weighting by reliability is analogous to assigning greater weight to more reliable information when updating a belief. Prediction error signals with high precision (inverse variance) have greater influence in updating the top-down predictions. Precision itself has to be inferred, both by the empirical variance in the sensory data itself, and according to prior expectations about precision. The optimisation of precision weighting, through updating of the precision expectations (precision-related priors), is frequently equated to attention within predictive processing (Clark, 2013; Feldman & Friston, 2010). Importantly for the current treatment, precision is thought to mediate both sensory attenuation - the top-down filtering out of afferent information, and affordances, where affordances refers to the latent possibilities for action given the capabilities of the agent (Cisek, 2007; Friston, Shiner, et al., 2012).

2.3 Control-oriented inference

In mandating that existence necessitates maintaining oneself within a limited repertoire of states via control-oriented predictive regulation (instrumental active inference) (Seth & Tsakiris, 2018), the FEP aligns itself with precursors of this view, cybernetic theories that build on control, feedback and predictive modelling (e.g., the “good regulator theorem”) (Conant & Ashby, 1970). Note that while a purely Helmholtzian view of the brain might cast it in terms of inferring hidden causes in the world, casting the predictive machinery in terms of being for ensuring continued existence means that the generative model is not constrained to veridicality. Rather than faithfully reconstructing the world, perception is “ultimately geared towards driving actions that preserve [the] physiological integrity of the organism. In other words, we do not perceive the world (and self) as it is, but as it is useful to do so” (Seth & Tsakiris, 2018, p. 975).

2.4 Homeostasis

Homeostasis refers to the tendency of living systems to keep an “internal balance” despite changes in the surrounding environment (Cannon, 1929). This has long been described in terms of control theoretic and cybernetic mechanisms, and more recently this homeostatic control is thought to involve interoceptive signals that report current physiological states (e.g., heart rate, or blood-bound glucose levels) (Craig, 2002; Damasio & Carvalho, 2013). One way to restore bodily conditions to favourable states is to engage autonomic reflexes – for example, a hyperthermic animal can perspire to cool down. Of course, autonomic regulation alone is not sufficient to ensure continued existence – to avoid hunger or thirst the animal must engage actions, such as seeking out food and water. Collectively, these actions are termed allostasis, the process via which the brain regulates the needs of the body (Corcoran & Hohwy, 2018; Corcoran, Pezzulo, & Hohwy, 2019; Schulkin & Sterling, 2019). Crucially, to stay viable on longer timescales, this action must be anticipatory – avoiding dyshomeostatic conditions before they arise (Pezzulo, Rigoli, & Friston, 2015; Sterling, 2012).

2.5 Active inference

The FEP regards homeostasis and allostasis as the central aspects of organic life, thus the autopoietic principles at the basis of the FEP act as a kind of “first prior” (Allen & Tsakiris, 2018). In other words, “[t]he brain is in the game of predicting the world, but only as a means to the end of embodied self-preservation” (Allen & Friston, 2018, p. 12). In so doing, the free energy principle collapses expected utility (instrumental value) and information gain (epistemic value) under a single quantity. On this approach, action planning is itself a form of inference, where preferences and goals are framed in terms of prior beliefs, such that these priors are fulfilled by action (Botvinick & Toussaint, 2012). Casting value and utility purely as inferential problems may at first appear unintuitive – if an agent finds itself in consistently adverse circumstances, then such adverse circumstances should, at first pass, seem to have high probability. However, “[t]he critical step in this logic is the assumption that evolution has equipped us with the belief that low utility states are low probability, due to the fact that if our ancestors spent a lot of time in those states they would be less likely to reproduce” (Gershman, 2019, p. 7). The so-called “first prior”, that of maintaining existence via homeostatic and allostatic regulatory behaviour, ensures that organisms seek to actively maintain internal and external conditions conducive to their own persistence.

Active inference refers to the process by which agents actively sample states of the world so as to reduce uncertainty and realise prior preferences, rendering the action selection process itself an inference problem. This arbitration occurs according to priors pertaining to expected free energy over a given course of action, or policy (Friston et al., 2018; Pezzulo et al., 2015). Expected free energy is the free energy an agent predicts itself to average in opting to pursue a particular course of action. Intuitively, some courses of action are more likely than others to lead to “expected” or desirable outcomes. A policy that has lower expected free energy is going to have a higher prior probability than a policy with higher expected free energy, because agents equipped with prior beliefs about their continued existence will pursue policies that reduce expected free energy (Friston et al., 2015, 2018; Kaplan & Friston, 2018). Crucially, agents engaging in active inference do not merely restrict themselves to the states they expect; rather they anticipate in order to minimise uncertainty about potential future outcomes (Friston et al., 2015, 2018; Pezzulo et al., 2016). This prospective form of control relies on the contextualization provided by higher levels in the inferential hierarchy, which anticipate the downstream consequences of actions and select policies accordingly (Friston, 2010; Pezzulo et al., 2015). Contextualisation here depends on the relative precision at various hierarchical levels, where “precision dynamics subsume the role of arbitration” (Pezzulo et al., 2015, p. 27). This approach bears similarities to other control-theoretic approaches, such as the affordance competition hypothesis (Cisek, 2007; Pezzulo & Cisek, 2016), where an affordance is a potential for action that avails itself to an organism in its action-oriented perception of environmental features.³ On this view, perceived affordances jostle for precedence and are arbitrated on the basis of the desirability of their predicted outcomes.

3.1 Temporal thickness

To successfully navigate the world over longer timescales, and select policies that result in survival – and not dispersion or non-existence – organisms must possess models of the future; in other words, they require deep temporal models (Friston et al., 2018). The generative models that endow organisms with the capability of inferring the consequences of future actions must have the property of temporal thickness, which allows the organism to anticipate the downstream consequences of potential actions, conferring the ability to select policies or action scripts that are favourable to the organism’s continued existence (Friston, 2018). The minimisation of surprise through active inference on the FEP involves acting so as to reduce uncertainty, and to do this the system must model itself across time and counterfactuals – that is, it must model what kind of agent it is at varying degrees of temporal depth. Self-modelling, then, emerges as a natural consequence of prospective action selection (Friston, 2018), where the principal function of a counterfactually rich self-model is to facilitate navigation of the affordance landscape and action selection across multiple interlocking timescales – for example expectations of what an agent can do on shorter timescales inform expectations of what the agent can do over longer timescales. The functional role of having a rich self-model, then, is that it enables the organism to predict outcomes across diverse policies, and endows the organism with “what if?” capabilities (Friston, 2018), which puts this picture into contact with mental time travel and offline simulation (Buckner & Carroll, 2007; Schacter, Addis, & Buckner, 2008).

3.2 Attuning to the world

Conceiving of the self-model through an active inference framework, a hierarchically deep self-model guides policy selection over various timescales in service of minimising expected free energy. In what follows, pain perception, viewed as arising through the violation of the prior of “healthy body condition” (Ongaro & Kaptchuk, 2019), will be used to illustrate how inferences about the self “attune” an organism to adaptive action opportunities.

One key advantage that the active inference account of self-modelling has over strictly Bayesian approaches is that it is goal-directed (Moutoussis, Fearon, El-Deredy, Dolan, & Friston, 2014). Classical models of pain perception as the consequence of physiological dysfunction are challenged by the efficacy of placebo treatments in relieving pain (Anchisi & Zanon, 2015), and cases in which pain is experienced without physiological disruption, as is often the case in chronic pain. Instead, there is evidence to suggest that affectively charged percepts, such as pain, are best understood as resting on the same inferential mechanisms as are assumed to underpin perception and action under a predictive processing framework (Büchel, Geuter, Sprenger, & Eippert, 2014). Bayesian models of pain perception (Morton, El-Deredy, Watson, & Jones, 2010) indicate that prior beliefs about the generation of painful percepts are integrated with current sensory data to infer the posterior or hidden worldly cause (the painful percept). Crucially, these pain percepts incorporate the “weight” or precision of past experiences when computing the current painful percept (Morton et al., 2010). On their own, however, these models of pain perception are silent on the functional role of pain as a motivator to an embodied organism (Moutoussis et al., 2014). Optimal inference about pain to the allostatically concerned organism is heavily dependent on the context, as anyone who has felt the pain of an injury only after danger is averted can attest to. In this way, pain perception is allostatically “tuned”: “organisms can tune their own pain perception according to both their prior beliefs and the specific biological goals they believe are attainable in that context” (Moutoussis et al., 2014, p. 70).

A Bayesian framework of pain perception, therefore, needs to represent the agency and aims of the organism. This is precisely what is afforded by conceiving of the self-model within an active inference framework (Friston, 2012b) – as this provides the necessary context to study the self-model, across multiple hierarchical levels. Like physical pain, and sharing the neural underpinnings of physical pain (Eisenberger, 2012), social pain is similarly understood in inferential terms, and does not scale with “damage” per se (for instance, social rejection), as evidenced by the wide range of sensitivity people have to the same physical manipulation (Eisenberger, Jarcho, Lieberman, & Naliboff, 2006). Accordingly, there is evidence to suggest that appropriately “tuning” emotional responses in social contexts allows for agents to approximate Bayesian inference in policy selection given bounded cognitive capacity and rationality. For example, on a “stag hunt” game,⁴ agents with “prosocial” preferences can outperform agents of similar cognitive sophistication that lack social biases (Yoshida, Dolan, & Friston, 2008).

3.3 Emotion

Conceptualising emotions in terms of a contextualisation of bodily states has historical roots dating back to the James-Lange theory of emotion (Cannon, 1927) and two-factor theory of emotion (Schachter & Singer, 1963). Lisa Feldman-Barrett has developed this approach specifically within the active inference framework as the “theory of constructed emotion” (Barrett, 2017). According to the theory of constructed emotion, emotions are constructed in the same manner as percepts, where priors are recruited according to context to make a “best guess” at the hidden causes of (interoceptive) sensory signals. On Barrett’s view, emotions arise through a context-sensitive inferential categorisation of interoceptive states. For this reason, emotions on this view are “constructions” – there are no neural or physiological signatures that reliably discriminate any emotional state (Siegel et al., 2018; Wager et al., 2015). Rather, physiological reactions in the body occur in order to prepare it for action, and these are categorised as emotions only contextually through the predictive models recruited to explain away the incoming afferent interoceptive signals. For example, heart rate increases or decreases depending only on an anticipated action – e.g., fight or flight – and given an emotional ascription only contextually – e.g., the same bodily state could be categorised as fear in one context and anger in another. Interoceptive inference is experienced as emotion in service of producing allostatic action (Barrett, Quigley, & Hamilton, 2016; Barrett & Simmons, 2015; Wilkinson, Deane, Nave, & Clark, 2019). In viewing the self-model in terms of hierarchical allostatic control, interoceptive inference on the hidden causes of bodily states pertaining to longer timescales tunes perception to the world and affordances differently, such that more abstract emotions might track regularities over longer time scales, informing policy selection thereon (Pezzulo, 2014), and allowing for more abstract and distal outcomes to be motivationally salient.

3.4 Hierarchically deep self-models

Viewing the self-model in terms of allostatic control renders selfhood fundamentally affective and action-oriented, such that different aspects of the self in a given context – precision on goals and preferences at different levels of the hierarchy – motivate behaviour and arbitrate between policies. On this view, the self-model inflects perception of possible actions in the world and mediates salience to facilitate the selection of policies with minimal expected free energy. Critical to this picture is the notion that these various models are associated with varying degrees of temporal depth (Pezzulo, Rigoli, & Friston, 2018). Deep generative models capture increasingly distal relations between actions and outcomes within hierarchical active inference, allowing for the coordination of behaviour across different hierarchical levels, enabling goals to become prioritised relative to current context (Pezzulo et al., 2018). The result is an inferential framework of hierarchically nested contextual complexity, in which lower levels track basic (and sometimes evolutionarily hard-wired) motivations or affordances, while higher levels track motivations and plans over deeper timescales. In this way, higher-level contextualization of lower sensorimotor functions optimises expected actions in terms of both long-range consequences of actions and anticipated future affordances. Goals at different levels of abstraction may, of course, conflict – for instance, resolving proximal interoceptive prediction error by eating chocolate cake might conflict with the longer-term goal of sticking to a diet (Pezzulo et al., 2018). Alternatively, temporary deviation from homeostatic set points at lower levels may be elicited to maintain higher level set points – such as a temporary change in blood pressure and adrenaline levels to engage fight-flight behaviour, with the goal of reaching safety and maintaining physiological integrity on a longer timescale. On the view of self-modelling in terms of allostatic control described, dimensions of the self at higher-levels constrain the self at lower-levels in that the self-model “actively shapes itself over time to align with those higher level regularities” (Hohwy & Michael, 2017, p. 370), for example long-term goals can be decomposed into intermediate short term-goals.

This section has explored how the self-model arises as a consequence of a system engaged in temporally deep active inference, as prior probabilities over particular policies depend on knowledge about what and where the system finds itself, and what actions are available to it (Friston et al., 2013; Moutoussis et al., 2014). Through active inference, agents can use their self-model to inform their goal and policy selection in order to arrive at high probability outcomes. This could entail assigning low probability to the self occupying states that are aversive, either physically or socially – with different hierarchical levels of the self-model contributing to different goal states. In this way, the hierarchical self-model determines salience – where “salience is literally defined by whatever has the most (or least) impact on visceral and autonomic homeostasis” (Allen & Tsakiris, 2018, p. 7), at increasingly deep spatiotemporal scales and levels of abstraction.

4.1 Psychedelic therapy

Recent years have seen a resurgence of interest in the therapeutic potential of psychedelics. Several studies have found preliminary evidence that with administration in controlled circumstances psychedelics can be both safe and therapeutic, with an emphasis on the importance of context in achieving therapeutic outcomes (Carhart-Harris et al., 2018). Interestingly, the positive therapeutic effects seem to scale with “peak” or mystical experience in the psychedelic state (Roseman, Nutt, & Carhart-Harris, 2018). Psychedelics have been shown to be effective in treating depression (Lyons & Carhart-Harris, 2018; Palhano-Fontes et al., 2019), obsessive-compulsive disorder (Moreno, Wiegand, Taitano, & Delgado, 2006), end of life existential distress (Griffiths et al., 2016), and have even been proposed as a potential treatment for disorders of consciousness such as the vegetative state and the minimally conscious state (Scott & Carhart-Harris, 2019). Carhart-Harris interprets the therapeutic effects as a result of “relaxing” high-level beliefs, allowing for a revision of pathological beliefs that have become overly dominant and resistant to revision, coined the “TIghtened BEliefs in Response to uncertainty” (TIBER) model (Carhart-Harris, 2019). The basic tenet is that under conditions of uncertainty the model falls back on “tightened” belief structures as a defence mechanism against intolerable stress and uncertainty. This fits with what we might expect under the FEP, as adopting shallow policies (such as addictive behaviours) may appear adaptive in the short term, rather than risking policies with greater expected free energy due to low precision or uncertainty. It has recently been proposed that psychedelics “relax” high-level priors in the generative model, allowing for the (context-dependent) revision of pathological high-level beliefs (Carhart-Harris & Friston, 2019). Both psychological insight and peak-experience in the psychedelic state appear to be predictors of long-term positive prognoses (Roseman et al., 2018).

4.2 Psychedelics in the predictive brain

The REBUS – “RElaxed Beliefs Under pSychedelics” – model of psychedelic function, offers a preliminary but promising model of psychedelic action where psychedelics, through 5-HT_2A agonism, “relax” high-level priors or beliefs (Carhart-Harris & Friston, 2019). Here, the focus will be on how this mechanism may be cast under the hierarchical predictive processing framework as modulating precision-weighting. To bring this into focus, this section will review how precision-weighting sets a variable Bayesian learning rate in order to highlight certain features relevant to understanding the effects of psychedelics within this framework. Christoph Mathys and colleagues have recently developed a mathematical tool for modelling Bayesian inference modulated by expectations of volatility known as the hierarchical Gaussian filter (Mathys, Daunizeau, Friston, & Stephan, 2011; Mathys et al., 2014). The hierarchical Gaussian filter Mathys posits allows a system to optimally balance the influence of prediction errors in changing environments – in other words, to adjust its learning rate.

4.3 Bayesian learning rate

To recap, the predictive processing framework asserts that the brain instantiates “generative models” of the causes of incoming sensory data, iteratively updating these predictive models in light of incoming “prediction error” (Clark, 2013). This predictive inference is thought to occur across a hierarchy of inferred causes, where high levels track causes and regularities operating over deeper spatial and temporal scales, and lower levels track regularities over shallower spatial and temporal scales (Kiebel et al., 2008). The picture of the living or cognitive system as one which needs to optimise its own learning rate emerges out of the operationalization of Bayesian inference in predictive processing, namely in terms of predictions and precision-weighted prediction errors (Mathys et al., 2014). According to predictive processing the prediction is given by the prior probability (which itself comes from the previous posterior) and the prediction error is given by the difference between the prediction and the incoming sensory evidence. Prediction error is weighted according to the relative precisions of the prior and the prediction error (where precision is equivalent to the inverse variance of each probability distribution). Intuitively, highly precise prediction error will drag the posterior closer to the distribution of the sensory evidence and further from the prior, and in cases of low precision weighting of the prediction error, the inference relies more on the prior. This determines the learning rate:

The more certain we are that the prior hypothesis is correct, the less we should be influenced by the prediction error (the evidence), which means that the learning rate is low. Conversely, the better the precision on the prediction error, the higher the learning rate; that is, the more we trust the quality of the evidence the more we should learn from it (Hohwy, 2017, p. 76)

In other words, the lower the learning rate, the greater the influence of top-down modulation from priors; the higher the learning rate, the greater the influence of the sensory evidence on the resulting posterior. Here, precision-weighting is the key mechanism – heavily weighted prediction errors drive a higher learning rate. In order to approximate Bayesian inference over time, it is essential for sensory systems to balance the learning rate appropriately. Over-reliance on priors means the system will fail to learn from experience, whereas over-reliance on sensory evidence (which may be noisy) will lead the system to “overfit”. On this picture, Bayesian perceptual inference that minimises prediction error on the appropriate timescale – that is, not overfitting or underfitting the model – needs to have a means of regulating the learning rate (Mathys et al., 2014). This is implemented by building models of precision, or expected uncertainty, where higher-level priors track longer-term regularities that inform the relative precisions of more basic priors (Hohwy, 2017). Optimising the learning rate, and in-so-doing minimising prediction error over time, is a critical challenge the brain faces. This is equivalent to selecting a time frame over which to minimise prediction error. Minimising prediction error over too short a timescale – overfitting – runs the risk of increasing prediction error in the long run. Conversely, failing to accommodate new evidence will lead to underfitting, a failure to update predictions in light of new sensory evidence.

4.4 Psychedelic action as high Bayesian learning rate

In line with the REBUS model (Carhart-Harris & Friston, 2019), the relaxing of high-level priors under classical (serotonergic) psychedelics⁵ means the system adopts a very high Bayesian learning rate – that is, it is in a highly plastic state, in accordance with research showing an increase in plasticity under psychedelics (Ly et al., 2018). This picture casts the perceptual effects of psychedelics – “tripping” – as rampant overfitting of the sensory data, resulting from a loss of the usual constraint exerted by higher-levels on lower-levels of the inferential hierarchy. This “rampant overfitting”, resulting from diminished influence from contextualising high-level priors tracking regularities on longer timescales, means the model fits a very short temporal scale, rapidly cycling through candidate models to account for the incoming sensory signal. It is worth highlighting a compatibility of the high Bayesian learning rate approach with other accounts of the mechanism of action of psychedelics in the predictive brain. The REBUS model posits the mechanism of action of psychedelics as reduced precision at high levels rather than increased precision at the sensory peripheries, as psychedelics appear to disrupt functioning via stimulation of 5-HT_2A receptors on deep pyramidal neurons, thought to encode high level priors or beliefs (Beliveau et al., 2017; Carhart-Harris & Friston, 2019; Jakab & Goldman-Rakic, 1998). In contrast, Philip Corlett and colleagues have suggested that psychedelics preserve normal priors and act by increasing sensory noise through enhanced AMPA signalling (Corlett, Frith, & Fletcher, 2009; Corlett, Honey, & Fletcher, 2016). On this approach, if the relaxation of high-level priors is indeed an effect of psychedelics, it could be understood to be the result of the fact that “the persistence and strength of the sensory signal suggest that there is something to be explained” (Corlett et al., 2009, p. 521). Arbitrating between these two mechanistic accounts and disentangling causation – whether the relaxation of high-level priors causes the reduction in sensory gating, or reduction in sensory gating eventually lowers precision at high levels – becomes very difficult here, and it is not clear a simplistic causal account is the right approach. While identifying the mechanisms of action is a key empirical and theoretical project, one potential advantage of the high Bayesian learning rate hypothesis is that it doesn’t distinguish between high precision at low levels and low precision at high levels, and as such remains agnostic over the mechanism of action while preserving the useful theoretical features of both accounts that will inform the theoretical account of ego-dissolution that follows.

4.5 Evidence for the high Bayesian learning rate hypothesis

A high Bayesian learning rate is concordant with the enhanced neural plasticity observed in individuals in a psychedelic state (Barre et al., 2016; Berthoux, Barre, Bockaert, Marin, & Bécamel, 2018; Ly et al., 2018). While an impairment to high-level cognition is found under psychedelics (Bayne & Carter, 2018), in line with the high Bayesian learning rate hypothesis, low-level learning (including extinction learning) and processing appears to be unaffected or enhanced in the psychedelic state (Carhart-Harris & Nutt, 2017; King, Martin, & Seymour, 1972; Romano et al., 2010). Further evidence for a high Bayesian learning rate under psychedelics is provided by a study looking at the effect of psilocybin on Kanisza triangles – perceptual objects where the brain “fills in” illusory contours using prior expectations – which found reduced filling in and a reduction in the related evoked potentials (Kometer, Cahn, Andel, Carter, & Vollenweider, 2011), concordant with the fact that a high Bayesian learning rate will reduce the effect of sensory history on current perception. In binocular rivalry studies – where different images are presented to each eye simultaneously, and are typically experienced as switching from one percept to the other – reduced switch rates and increased likelihood of the percept being a fusion of the two images has been observed under psilocybin (Carter et al., 2007, 2005), suggestive of less influence of priors on constraining current perception. Oddball paradigms are also suggestive of a weakened influence of priors on perception under psychedelics. In a sequence of tones, an “oddball” tone (unexpected given prior experience and context) generates a “mismatch negativity”, an evoked brain response which has been interpreted in predictive coding terms as prediction error violating the expectations of the sequence (Garrido, Kilner, Stephan, & Friston, 2009). Under LSD, the surprise response to oddball stimuli is blunted, suggestive of a weakened influence of prior expectations (Timmermann et al., 2018).

Arguably, there is also phenomenological evidence for the high Bayesian learning rate hypothesis. Perhaps most eloquently articulated by Aldous Huxley: “Visual impressions are greatly intensified and the eye recovers some of the perceptual innocence of childhood, when the sensum was not immediately and automatically subordinated to the concept” (Huxley, 2010, p. 12).

This observation lends itself to a straightforward translation into the terms of predictive processing, where “subordinated to the concept” can be understood as “constrained by higher-level priors”. More generally, psychedelic phenomenology such as dynamic distortions of spatial dimensions, where things change dramatically in size and shape can be understood as a failure of high-level priors to canalise and constrain lower level predictions.

5.1 Ecstatic ego-dissolution and challenging experiences

The question remains as to why the hypothesised collapse in the temporal thickness of the self-model under psychedelics can be both ecstatic and of enduring therapeutic value. To bring this into focus, it’s worth recapitulating core features of the self-model provided earlier. Recall, interoceptive inference on states of the embodied self “attunes” organisms to their affordance landscape, where inferences about the state of the embodied self (e.g. hunger) prescribe certain prediction error minimising policies (e.g. finding food). Inferences pertaining to allostatic consequences on longer timescales may mean higher-level imperatives trump lower-level drives, such as choosing to abstain from chocolate cake to stay healthy (Pezzulo et al., 2018). In the case of basic bodily needs, as described, these variables are controlled (Seth, 2015) through action – active inference is deployed to bring the world into line with predictions, rather than adjusting predictions (via perceptual inference) to conform to the world – for instance eating when hungry (Pezzulo et al., 2015). In just the same way that a hungry organism can act so as to harvest confirmatory evidence for the hypothesis “I am sated”, hypotheses relating to higher-levels of the self-model geared towards control of outcomes on longer timescales act to constrain action in the present to bring downstream outcomes closer in line with the prior expectation. Overly precise priors driving action on a long timescale which are failing to be fulfilled, on this view, would be a persistent cause of suffering, due to the system consistently failing to meet (or align actions towards) the goal state (Hesp, Smith, Allen, Friston, & Ramstead, 2019). Under the model of psychedelic-induced ego-dissolution proposed, the high-precision high-level priors geared towards control on multiple timescales cease to exert influence on the system due to the proposed lowering of precision of high-level priors under psychedelics. If action ordinarily arises from a process of minimising deviations between the organism’s actual (inferred) and desired trajectory (Friston et al., 2010), the loss of precision on high-level priors means that, instead of driving action policies, they lose influence on the rest of the system and cease to structure pre-reflective self-consciousness to orient to action opportunities favouring their fulfilment. As these prior beliefs are relaxed, they instead become amenable to perceptual revision from the influx of (highly precise) interoceptive and exteroceptive information. The collapse in temporal depth in the psychedelic state is therefore not experienced as a loss of allostatic control, precisely for the reason that the priors pertaining to longer timescales are no longer asserting an influence on the system and constraining action (and perception) in their usual manner. This picture seems to align well with phenomenological reports of ego-dissolution: “It felt as if ‘I’ did no longer exist. There was purely my sensory perception of my environment, but sensory input was not translated into needs, feelings, or acting by ‘me’” (unpublished online survey data quoted in Millière et al., 2018, p. 7). Peak experiences under psychedelics, then, could be understood as absence of prediction errors relating to allostasis due to a flattening of the temporal depth of the affordance landscape, resulting in the feeling of “oceanic boundlessness” – a sense of immense well-being and peace. Here, the “itinerant strategies” to stay within our “species-specific window of viability” (Clark, 2013, p. 13), are no longer necessary as the “first prior” – the expectation or imperative for existence – is being met without conditions.

Following the TIBER model, many psychopathologies may be due to high precision on high-level priors (Carhart-Harris, 2019; Clark, Watson, & Friston, 2018). Peak psychedelic experience may act as a “reset” allowing for revision of entrenched high-level beliefs that structure pre-reflective self-consciousness (and, accordingly, the affordance landscape) – opening up new domains of salience and possibility for meaningful engagement with the world, through revised and retuned self-models. Increased bottom-up information flow (particularly from the limbic system), through a high Bayesian learning rate, may make entrenched high-level priors amenable to revision via perceptual inference rather than driving control via active inference. This lays the theoretical groundwork for why psychedelics may effectively treat depression: if depression is underpinned by a high precision prior of low allostatic self-efficacy (Stephan et al., 2016), it follows that relaxation and revision of this prior should alleviate depressive symptoms. Finally, (and speculatively), if the account of “retuning” of self-models under psychedelics presented here generalises to the bodily self (which the experiential changes in bodily selfhood would suggest) this account is suggestive of a potential role for psychedelics in the treatment of chronic pain, and for phantom limb pain – for which there has already been promising results (Fanciullacci, Bene, Franchi, & Sicuteri, 1977; Ramachandran, Chunharas, Marcus, Furnish, & Lin, 2018).

The primary focus so far has been on “peak” experiences, due to the growing number of papers indicating they are central to positive long-term therapeutic outcomes (Roseman et al., 2018). However, while generally psychedelics are thought to be very low risk (Nutt, King, & Phillips, 2010), and there is evidence to suggest they are protective against mental health problems (Hendricks, Thorne, Clark, Coombs, & Johnson, 2015), acute and occasionally persistent adverse psychological reactions do sometimes occur (Strassman, 1984). While “complete” ego-dissolution is described as a “state of complete surrender, associated bliss, and union with all things” (Carhart-Harris & Friston, 2019, p. p321), “incomplete” ego-dissolution – due to psychological resistance or an insufficient dose – can be characterised by intense fear, anxiety, or distress. On the account presented in this paper, this can be understood as resulting from psychological resistance, where psychological resistance here may be conceptualised as a high-precision prior on being able to control the experience, that is maintained though fear-driven endogenous attention. Failure to control the experience, in violating the highly precise prior for the goal state of control, is then experienced as a loss of allostatic control, bringing with it feelings of intense fear or distress. In therapeutic contexts, encouraging users to “let go” and “surrender” to the experience (Richards, 2015), could be understood in these terms, as discouraging the user from putting high (endogenous) precision on a prior for control that could result in adverse experiences when unfulfilled. These considerations highlight the essential importance of context in achieving therapeutic outcomes (Carhart-Harris et al., 2018).