Consciousness in active inference: Deep self-models, other minds, and the challenge of psychedelic-induced ego-dissolution

This theory-building paper (2021) expands REBUS model to accommodate a theory of consciousness informed by an assessment of psychedelic induced ego-dissolution within the context of the active inference framework. The author contends that although these selfless states of consciousness are rare, they reveal normally congruent processes of computational and phenomenal self-modeling which demonstrate the principle that phenomenal models of the self ‘shape’ the subjectivity of our experiences.

Abstract

“Predictive processing approaches to brain function are increasingly delivering promise for illuminating the computational underpinnings of a wide range of phenomenological states. It remains unclear, however, whether predictive processing is equipped to accommodate a theory of consciousness itself. Furthermore, objectors have argued that without specification of the core computational mechanisms of consciousness, predictive processing is unable to inform the attribution of consciousness to other non-human (biological and artificial) systems. In this paper, I argue that an account of consciousness in the predictive brain is within reach via recent accounts of phenomenal self-modelling in the active inference framework. The central claim here is that phenomenal consciousness is underpinned by ‘subjective valuation’-a deep inference about the precision or ‘predictability’ of the self-evidencing (‘fitness-promoting’) outcomes of action. Based on this account, I argue that this approach can critically inform the distribution of experience in other systems, paying particular attention to the complex sensory attenuation mechanisms associated with deep self-models. I then consider an objection to the account: several recent papers argue that theories of consciousness that invoke self-consciousness as constitutive or necessary for consciousness are undermined by states (or traits) of ‘selflessness’; in particular the ‘totally selfless’ states of ego-dissolution occasioned by psychedelic drugs. Drawing on existing work that accounts for psychedelic-induced ego-dissolution in the active inference framework, I argue that these states do not threaten to undermine an active inference theory of consciousness. Instead, these accounts corroborate the view that subjective valuation is the constitutive facet of experience, and they highlight the potential of psychedelic research to inform consciousness science, computational psychiatry and computational phenomenology.”

Author: George Deane

Summary

Predictive processing approaches to brain function can illuminate the computational underpinnings of a wide range of phenomenological states, including consciousness. However, it remains unclear whether predictive processing is equipped to accommodate a theory of consciousness itself.

Introduction

Phenomenal consciousness has now been a serious scientific study for at least 30years. The predictive processing framework has generated considerable excitement for its potential contribution to consciousness science. A theory of consciousness within predictive processing remains elusive, largely due to the fact that predictive processing is not exclusively concerned with conscious processing. However, I aim to show that predictive processing has the resources to deliver a fully-fledged theory of consciousness via a theory of subjectivity grounded in self-modelling.

Active inference is a process theory of the free energy principle, and it can be applied to understand the adaptive behaviour of bacteria, plants, social and cultural dynamics, and natural selection.

The structure of the paper is as follows: I give an overview of the relevant mechanics of the active inference framework, I propose a self-modelling theory of subjectivity within active inference, and I consider how self-models shape ordinary self-consciousness. The selflessness challenge is an objection to subjectivity theories of consciousness. I argue that the selflessness challenge is best understood in terms of a model of allostatic control, and that this model accounts for how the system can still be conscious without the typical structure of experience provided by deep self-models.

The active inference framework

The active inference framework is a formalization of allostasis, a process by which organisms maintain homeostasis by keeping physiological states within reasonable bounds. This process is known as prospective control or ‘predictive regulation’.

The active inference framework formalizes allostasis in terms of a single imperative: to minimize the divergence between expected and observed outcomes under a generative model. This imperative is achieved through phylogenetically endowed high precision prior expectations.

A hierarchical generative model is specified in terms of probabilistic beliefs about how observations relate to the states of the world that cause them, beliefs about how the states evolve over time and prior beliefs. Inference is made tractable by optimization of a posterior although variational inference.

Active inference formalizes allostasis (and action) as an inference problem under the free energy principle, where the action with the highest prior probability minimizes expected free energy, and organisms rely on deep temporal models to select optimal action policies.

Action selection is understood in terms of Bayesian model selection, where possible action policies are scored with respect to the expected free energy associated with pursuing a given policy. This process is dubbed ‘allostatic control’, and it is underwritten by hierarchically deep inference about the precision on expected free energy.

Pragmatic value and prior preferences

In active inference, agents change the world to make it conform to prior preferences, rather than changing beliefs to conform to the world (i.e. perceptual inference). This is key in allostasis, where temporary deviations from a homeostatic setpoint are allowed in order to realize certain actions. Over the course of ontogeny, an organism will acquire increasingly deep temporal scale prior preferences, which determine the expected free energy of a given policy.

Epistemic value

Self-evidencing agents act to realize prior preferences, but also to realize epistemic value. This is accounted for within a formulation of epistemic action that considers curiosity and novelty-seeking behaviour.

Consciousness in active inference

In this section I argue that consciousness arises from hierarchically deep self-models that evaluate the fit of their model with the world across multiple timescales. This inference informs action and policy selection across multiple timescales.

I sequentially unpack the inferential architecture of the hierarchical generative model and argue that this architecture relates to conscious contents in four subsections: perception, precision, active inference, and affectivity. This inference about endogenous control shapes subjectivity.

Perception

The generative model is a way to think about perception, where the brain makes a best guess at the causes of sensory signals. It includes beliefs about the most likely state of the world prior to any observation.

The system needs to infer the probability of hidden states given the observations, and this is achieved through variational Bayes. This process is known as model inversion, and perception can be understood as posterior state estimation.

Precision

Predictive processing architectures benefit from radical contextual flexibility afforded by ‘precision-weighting’, where heavily weighted priors or prediction errors exert greater influence in determining the resulting posterior inference. Attentional processes operate on second-order statistics like precision.

Predictive processing involves integrating information from across modalities to infer the hidden causes of sensation. For instance, auditory, olfactory, tactile and interoceptive information have all been shown to influence visual experience.

Cue integration is one example of how integrating precision-weighted informational streams gives rise to the resultant percept. The reliability of the cues can be varied to make one cue more reliable than the other.

A gist perception of a scene engages past experience to generate a most likely prediction about an object’s identity. This prediction is fed back to early visual areas to speed perception by constraining the hypothesis space of possible interpretations.

The brain uses belief updating schemes to approximate Bayesian inference by utilizing priors and incoming sensory data to arrive at a posterior estimate. Prediction errors are a central concept in predictive processing and the free energy principle.

Predictive processing gives a compelling story about the contents of perception, where conscious perception is determined by the prediction or hypothesis with the highest overall posterior probability.

This section will look at how active inference can be used to understand phenomenal selfhood, and how ongoing policy selection can be understood to underpin agent selfhood.

The phenomenal self-model is understood as the content of the conscious self, including current bodily sensations, present emotional situation, plus all the contents of your phenomenally experienced cognitive processing. The self is thought to be a hypothesis or latent state that can be associated with a self-model.

Inference about the control of sensation via action is linked to the phenomenology of being an agent, and is temporally deep, because expectations of the consequences of actions are not confined to the immediate future, but can predict abstract and distal outcomes.

A central idea in this account is that the system must infer ‘itself’ as able to bring about the self-evidencing consequences of the action. This is done by lowering the precision of sensory evidence.

In active inference, physiological sensory attenuation is critical for movement initiation. Higher-level prior beliefs attenuate current sensory evidence and higher precision is afforded to the anticipated sensory consequences of the desired action. Motor control is based on a specification of the desired sensorimotor endpoint, and on perceptual control theory. Physiological sensory attenuation aids in entertaining counterfactual hypotheses about oneself, in order to generate the self-fulfilling prophecy of moving.

Self-attenuation mechanisms are thought to underpin various states of altered self-experience, such as the rubber hand illusion.

The comparator model posits that the sense of agency is formed by comparing the sensory consequences of an action with the intended consequences of an action. A mismatch between motor output and sensory input justifies the attribution of sensory outcomes to exogenous causes.

The system filters out irrelevant inputs by attenuating precision on sensory inputs. This is thought to be the cause of inability to tickle oneself.

Affective inference

Affective inference is an inference about allostatic control that acts to ‘tune’ the organism to possibilities for self-evidencing action in the environment. Pain perception is a great example of ‘tuning’ to the current context, and is influenced by a host of contextual factors. For example, attention, expectation, conditioned pain modulation and placebo responses can all have a profound influence on experience.

Inference about endogenous control of self-evidence outcomes can be understood as an inference about ‘subjective fitness’, which is the expected precision of the organism’s phenotype-congruent action model. This is essential for organisms to persist and perform adaptive actions in volatile environments.

Inference about the reliability of the action model allows the system to increase or decrease precision on the current policy. This is done by reimagining affective valence within the active inference framework.

A sensitivity to worse than expected rates of prediction error reduction over time drives the system to switch to more tractable goals, such as a task with a better expected rate of prediction error reduction.

The sense of being a self is inherent in active inference, and is underpinned by estimation of the precision of its own action model. This subjective valuation of the precision of the action model spans multiple levels of the hierarchy, and can be out of step with reality.

The mechanisms underpinning phenomenal consciousness on the current picture act to tune the organism to adaptive action in the world across multiple interlocking timescales, and even state estimation associated with perceptual inference is determined by the overarching inference about control of self-evidenced outcomes.

A conscious agent encounters a structured world through an interoceptivelymediated sense of mattering, and makes inferences about the fit between the actual and expected outcomes of actions. This inference is understood as affective in virtue of being enslaved by higher-level goals.

Interpreting sensory input as meaning for the action model allows the system to arbitrate between competing affordances on different timescales, and this connects the current story to numerous accounts of phenomenal consciousness.

The shape of subjectivity: disruptions in (self-)consciousness

Depersonalization disorder is understood as an inferred loss of allostatic control, whereby the system posits itself as causally inefficacious at realizing self-evidencing outcomes across contexts. Major depression is similarly understood as an inferred loss of allostatic control.

This phenomenology is contrasted with a perceived gain in allostatic control in meditation practitioners, as precision on prior preferences becomes increasingly under endogenous control. This is because focused attention meditation involves the endogenous withdrawal of precision from prior preferences.

Consciousness in other systems

This section will describe how the neural mechanisms associated with the allostatic control model enable organisms to act adaptively in their environment, enabling both adaptive motor control and determining the perceptual salience in the given context.

Holst and Mittelstaedt (1950) identified an interpretative problem as to whether sensory signals arise from the environment or the animal’s own muscles and movement. This problem is solved by corollary discharge, which is a top-down prediction of sensory consequences of actions that suppress reafferent inputs.

Crapse and Sommer (2008a) distinguish between lower-order (reflex inhibition and sensory filtration) and higher-order (sensory analysis and sensorimotor learning/planning) corollary discharge. The nematode Caenorhabditis elegans has a simple nervous system with 302 neurons and uses lower-order corollary discharge to inhibit reflexes that would be triggered by reafference. However, there is no evidence that this mechanism contributes to a structured model of the self or a model of action-outcome contingencies.

Higher-order corollary discharge is involved in predictive control in perceptual cohesion and action sequencing in bats, primates and juvenile songbirds. This process involves continuous updating of an internal record of current state.

Non-human animals’ complex and context-sensitive sensory filtration mechanisms may be the most promising hallmarks of consciousness. ‘Unlimited associative learning’ may also be a marker of the evolutionary transition to minimal consciousness.

The selflessness challenge

Many approaches to understanding consciousness have argued that self-consciousness is necessary or constitutive of consciousness itself. For instance, phenomenological theories have argued that self-consciousness is an integral and constitutive feature of phenomenal consciousness.

Milliere and Metzinger (2020) argue that subjectivity theories are committed to the necessity claim that self-consciousness is necessary for consciousness, and that the active inference framework provides a unifying theoretical framework for understanding a host of closely related phenomena.

The claim that self-consciousness is necessary for consciousness may present a problem for active inference in delivering a theory of consciousness, because experiences of altered selfhood present a problem for theories of consciousness that claim self-consciousness is necessary for consciousness.

The subjectivity theorist has two options: (1) deny that partially selfless states of consciousness truly present a challenge to subjectivity theories, or (2) deny that there really are states of consciousness lacking in self-consciousness.

Serotonergic psychedelics such as 5-methoxy-DMT (5-MeO-DMT) induce profound alterations in phenomenology, including profound alterations in self-consciousness. These alterations are the strongest evidence against the view that self-consciousness is necessary or constitutive of consciousness.

These experiences provide counterevidence to the necessity claim that self-consciousness is necessary for consciousness. Do these experiences pose a problem for the active inference account of consciousness?

Psychedelics and selflessness in active inference

The REBUS model proposes that psychedelics relax the precision of high-level priors, thereby liberating bottom-up information flow, and that this leads to instability in higher-level representations, which is manifest as the phenomenological effects of psychedelics.

The system adopts a high Bayesian learning rate on sensory evidence, which results in a greater precision on sensory evidence and less constraint imposed by higher-level priors. This is crucial for the system to approximate Bayesian inference over time.

Psychedelic experiences are associated with positive emotions such as exhilarated elation with unmotivated laughter, deep feelings of peace, exuberant joy, and hedonistic pleasure. These emotions are due to the high precision on sensory evidence.

Psychedelic-induced ego-dissolution

Deane (2020) describes psychedelic-induced ego-dissolution as resulting from a failure of sensory attenuation. This failure can be explained by a failure to attribute self-generated outcomes to endogenous rather than exogenous causes, such as a perceived loss of agency over one’s thoughts.

Corollary discharges act to cancel out self-generated sensory outcomes via sensory attenuation, and unexpected consequences are then attributed to exogenous rather than endogenous causes. This leads to ego-dissolution, which is experienced as a global dissolution of selfhood.

Affective tone

While ego-dissolution is described as being devoid of self-consciousness, it is nonetheless described as a highly conscious state, characterized by affective extremes.

On the current account, complete ego-dissolution is underpinned by two closely related computational mechanisms: pragmatic value and prior preferences. The relaxation of the constraining influence of high-level priors means they cease to structure consciousness to engage the organism in their fulfilment, and as such end their associated suffering. Deane (2020) highlights that the lessened influence of prior preferences accords with descriptions of the phenomenology of ego-dissolution.

‘Complete’ ego-dissolution may be characteristically ecstatic because the system infers that the current state is realizing great epistemic value, and the relaxation of high-level priors is associated with increased sensory precision.

Responding to the selflessness challenge

Psychedelic-induced ego-dissolution threatens active inference theories of consciousness grounded in self-modelling, because subjectivity is underpinned by an inference about allostatic control, which evaluates how well the system is bringing about expected outcomes, informing policy selection, and shaping subjectivity.

Selfless states such as those occasioned by psychedelics threaten the necessity relationship between phenomenal self-modelling and consciousness, but the inferential process evaluating the fit between model and world remains intact.

The present account puts subjective valuation as the most basic constitutive feature of a conscious experience, and psychedelic-induced ego-dissolution brings this view into sharp relief. Affective valence can be understood as the most fundamental part of conscious experience.

On this view, feelings are the most basic constitutive phenomenal states, and they guide the organism to fitness-promoting states.

Many will not be satisfied with this characterization of self-consciousness, and it may be that the selflessness challenge is best addressed by a more stringent definition of self-consciousness, in which the affective inference present in states of drug-induced ego-dissolution is deemed not to qualify as self-consciousness.

The current account does not rule out the possibility of states of consciousness that vary on these dimensions, such as states of consciousness that score low on domain-general affectivity and high on sensorimotor consciousness.

Godfrey-Smith notes that some spiders demonstrate complex perceptual capacities, but score low in respect to evidence for complex or varying motivational states. However, on the current account, both sensorimotor and evaluative aspects of consciousness are underpinned by a common evaluative inference about realizing phenotype-congruent outcomes across multiple timescales.

The active inference approach to consciousness and subjectivity put forward in this paper can accommodate phenomenological states largely lacking in self-consciousness, and illuminate the core aspect of phenomenal self-modelling – subjective evaluation.

Conclusion

This paper argues that phenomenal consciousness is best understood within predictive processing in terms of the deep self-models inherent in the active inference framework. It also argues that psychedelic-induced ego-dissolution is not troubling for an active inference theory of consciousness.

PDF of Consciousness in active inference: Deep self-models, other minds, and the challenge of psychedelic-induced ego-dissolution