Perception is in the Details: A Predictive Coding Account of the Psychedelic Phenomenon

This theory-building paper (2017) proposes that the breakdown of top-down prediction by psychedelics happens through them making serotonin (5-HT) 2a receptors (in layer V pyramidal neurons) hyperactive.

Abstract

“Psychedelic substances are used for clinical applications (e.g., treatment of addictions, anxiety and depression) as well as an investigative tool in neuroscientific research. Recently it has been proposed that the psychedelic phenomenon stems from the brain reaching an increased entropic state. In this paper, we use the predictive coding framework to formalize the idea of an entropic brain. We propose that the increased entropic state is created when top-down predictions in affected brain areas break up and decompose into many more overly detailed predictions due to hyper activation of 5-HT2A receptors in layer V pyramidal neurons. We demonstrate that this novel, unified theoretical account can explain the various and sometimes contradictory effects of psychedelics such as hallucination, heightened sensory input, synesthesia, increased trait of openness, ‘ego death’ and time dilation by up-regulation of a variety of mechanisms the brain can use to minimize prediction under the constraint of decomposed prediction.”

Authors: Sarit Pink-Hashkes, Iris van Rooij & Johan Kwisthout

Notes

This paper builds further on the entropic brain hypothesis by Carhart-Harris and colleagues (2014).

“They suggested that this higher variance of activity allows for enhancement of the repertoire of possible states over time, and introduced the term Entropic Brain to describe this higher entropic state.”

“We propose that the increased entropic state is created when top-down predictions in affected brain areas break up and decompose into many more overly detailed predictions due to hyper activation of 5-HT2A receptors in layer V pyramidal neurons.”

“According to predictive coding, perception is a continuous process of combining the brain’s previous knowledge with new incoming data by using Bayesian updating, so as to best represent the environmental causes of its sensory input. This enables the brain to predict its sensory inputs.”

“Muthukumaraswamy et al. (2013) found, following administration of Psilocybin, a desynchronization of neural activity especially in the slower alpha and beta rhythms, meaning neurons were acting in a more disjoint and separate way, suggesting that the brain was at a higher entropic state”

”In most cases this will result in higher prediction error from lower layers as these decomposed predictions ‘explain away’ less of the prediction error from lower layers than normal. The ‘extra’ predictions being activated are likely to be dependent on a subject’s personal experiences and history. In general we should expect a flattening of the prediction distribution, and well-established prediction categories that contain many subcategories will be affected more than predictions with fewer subcategories.”

“The noisier the bottom-up signal, the more the top-down predictions influence perception (Seth, 2014). Under decomposed predictions, lowering precision of sensory data can result in misclassification of the data. The brain’s best explanation for the imprecise ‘noisy’ data might be one of the sub-threshold predictions that got activated. This will result in a ‘hallucination’. Psychedelics are known to both obscure and distort perceptual data as well as add clarity and give the sense of enhanced resolution. These two different sides of the psychedelic state are dependent on the precision of the bottom-up data, i.e., the noisiness of the setting.”

“Predictions from other layers of the brain hierarchy that were not affected by activation of the 5-HT2A receptors can be upregulated by either increasing their relative strength or lowering their level of detail. This will cause the predictions from these layers to enforce their predictions on more of the incoming data. Google’s deep neural network ‘deep dream’1 (originally created for identifying images) illustrates how this might happen. By allowing different layers of the network to strengthen their predictions these networks were able to produce hallucinatory effects.”

“By creating motor output, for instance while walking or dancing, the mechanism of active inference (in which motor output minimizes proprioceptive prediction error between the expected and actual position of one’s limb, bringing the actual position closer to the expected position; see, e.g., Brown et al., 2011) might enable the brain to lower prediction errors stemming from other parts of the brain too.”

About experienced psychonauts not having that much of a trip.

“This might happen as a result of the brain’s attempt to minimize prediction error by lowering the weight of the prediction error or attributing this higher prediction error to ‘inherent’ noise that does not need to be explained.”

And “The brain could learn that this state is inherently noisier and lower the weight of the prediction error. We can only postulate that this might happen through affecting the dopamine system which has been implicated in precision weighting of prediction error (Friston et al., 2012).”

Learning over the long term (for disorders)

“Using the predictive coding framework, depression, addiction and obsessive compulsive disorders have been suggested to stem from overly strong and narrow predictions from certain networks that get ‘stuck’ (Edwards et al., 2012) and aren’t updated based on the bottom-up data. Momentarily decomposing these predictions by 5-HT2A agonists, especially with a combination of supportive bottom-up information coming from a therapeutic setting, might lead to long term model updates.”

And “A long term model update that psychedelic are known to cause is increasing the trait of ‘openness’ (MacLean, Johnson, & Griffiths, 2011). The mechanism we suggest to explain this is as follows. A higher prediction error state caused by administration of 5-HT2A agonists coupled with a positive rewarding setting, leads to surprise becoming a more sought after state. Interest in exploring the unknown and trying new things might grow and people might be ‘motivated to enlarge their experience into novel territory’ which is what defines the trait of openness (DeYoung et al., 2009).”

Time

“A few minutes can subjectively be perceived as taking much longer. Here we postulate that subjective sensation of time is dependent on the amount of prediction error and possibly prediction updates the brain makes in order to minimize prediction error.” … “Ulrich (2006) discovered that the extent to which a stimulus can be predicted affects time perception, with unexpected 2 This is known as the N170 event-related potential (ERP). stimuli perceived as longer. Similarly, Tse et al. (2004) found that a stimulus that stands out as different from all the others in a series appears to last longer than the other stimuli. An increase of prediction updates might cause the subjective feeling that more time has passed. This is similar to the common feeling that the first day of a journey to another country seems longer because it is filled with so many new experiences and so many prediction updates must happen in that day.”

Ego Death

“Apps & Tsakiris (2013) describe a predictive coding account of the neural and computational basis of self-recognition. Here, one’s body is recognized as the most likely “me”. This probabilistic inference arises through the integration of information from hierarchically organized unimodal systems in higher-level multimodal areas. As we have seen, the brain’s attempt to minimize increased prediction error induced by psychedelics breaks down this hierarchical structure which might lead to a total inability to distinguish between environment and self and the unique perception of ‘oneness’ described by many experiencing ‘ego loss’. While Apps & Tsakiris’ account deals with the ‘minimal self’, we postulate looking at the ‘higher ego’ as a collection of high-level relatively inflexible predictions regarding the future behaviour of the ‘self-organism’ in a variety of situations. Following administration of 5-HT2A agonists these predictions will break up based on the subjective pieces of information compromising this category. This relaxation of otherwise rigid predictions about the self might explain positive results for treatment of depression and addiction after administration of psychedelics that have been reported (Nichols, 2016).”

Summary

Psychedelic substances are used for clinical applications as well as investigative tools in neuroscientific research. We propose that the increased entropic state is created when top-down predictions in affected brain areas break up and decompose into many more overly detailed predictions.

Introduction

Psychedelics cause their effects by being partial agonists of serotonin receptors, with particular importance being given to those expressed on apical dendrites of neocortical pyramidal cells in layer V.

In this paper, we propose a predictive coding account of the effect of psychedelics that explains the various cognitive effects of psychedelics by up-regulation of a variety of mechanisms the brain can use to minimize prediction under the constraint of decomposed predictions.

In the next section we will introduce the predictive coding account and show how it can explain the cognitive effects of psychedelics.

A Predictive Coding Primer

Aldous Huxley proposed that perception is a door between things that are known and things that are unknown. Predictive coding is a contemporary account of brain processing that uses Bayesian updating to combine previous knowledge with new incoming data to best represent the environmental causes of its sensory input.

Recently, Kwisthout and colleagues proposed that more detailed predictions cause higher prediction errors. This work is based on the idea that higher cognitive functions are better described by categorical probability distributions rather than the traditional Gaussian densities.

Bastos et al. (2012) suggest that superficial layers of cortex show neuronal synchronization and spike-field coherence predominantly in the gamma frequencies, while deep layers prefer lower (alpha or beta) frequencies.

Psilocybin administration leads to a desynchronization of neural activity, suggesting that the brain is at a higher entropic state. This desynchronization is likely triggered by 5-HT2A receptor-mediated excitation of deep pyramidal cells.

A Predictive Coding Account of the Psychedelic State

Psychedelics affect the brain by activating 5-HT2A receptors in layer V pyramidal cells, which lowers the threshold of individual neuronal firing and thus desynchronizes the activity of the neuronal population.

Based on Kwisthout et al.’s (2017) notion of state space granularity in predictions, we suggest that hyperactivation of cells in layer V decomposes the broad categorical prediction into sub categories, creating a set of higher detailed predictions. These higher detailed predictions now dominate perception.

A person walking in the forest receiving sensory input will make more predictions when under influence of psychedelics, and these more detailed predictions will bring about a higher entropy state, which will result in higher prediction error from lower layers.

The importance of bottom-up data in this process

A clear perception of an environment can be achieved when precise environmental data combines with decomposed higher detailed predictions. However, due to environmental changes and noise, this clear perception is not likely to stay stable over time.

Prediction error minimization and the psychedelic state

Under normal conditions, the brain can decrease prediction error in several ways, including updating predictions, active inference, changing the weight of predictions, and long-term learning effects.

Updating the predictions

When predictions are decomposed, a smaller amount of sensory inputs will be explained by any specific prediction, which will cause increased prediction errors. This will cause a destabilization of perception, where objects, scenes and even abstract thoughts will morph.

Predictions from higher layers of the brain hierarchy can be upregulated by activation of the 5-HT 2A receptors. This causes predictions from these higher layers to enforce their predictions on more of the incoming data, and this is likely what causes visual hallucinations.

Acting on the Environment

The mechanism of active inference, in which the brain creates motor output, minimizes proprioceptive prediction error, and thus decreases uncertainty. This mechanism can help explain why hallucinations seem to grow stronger while sitting still.

Changing Weight of the Prediction Error

Psychedelic drugs can cause tolerance in the brain, where the first few experiences feel stronger than later experiences and increased dosage is needed to reach the same state. This might be due to the brain minimizing prediction error through the dopamine system.

Long Term Learning Effects

Within the predictive coding framework, changes in the network connectivity lead to long term learning. 5-HT2A agonists have been shown to enhance associative learning in rabbits, and this might explain the success of recent clinical trials that have used 5-HT2A agonists to treat depression, addiction and obsessive compulsive disorders.

Psychedelics cause a long term model update that increases the trait of openness. This is explained by a higher prediction error state caused by administration of 5-HT 2A agonists.

Psychedelics research findings reinterpreted

Kometer et al. (2006) presented Kanizsa triangles to subjects after administration of psilocybin and found that the ERP was weaker than normal, suggesting that the predictions were decomposed.

Spitzer et al. (1996) found that Psilocybin increased indirect semantic priming, which suggests that Psilocybin may bring cognitive contents to mind that under normal circumstances remain nonactivated.

A well-documented effect known as ‘Time Dilation’ occurs when the brain makes more prediction updates in order to minimize prediction error. This causes the subjective feeling that more time has passed.

The last phenomenon we would like to touch upon is the notion of ‘Ego death’ many psychedelic users report. This phenomenon might be explained by the brain’s attempt to minimize increased prediction error induced by psychedelics and the unique perception of ‘oneness’ described by many experiencing ‘ego loss’.

Conclusions

In this paper we proposed a computational theory explaining the effects of psychedelics in terms of the predictive coding account of cortical processes. This theory further explicates the Entropic Brain hypothesis in terms of predictive coding, and suggests that serotonin might have a role in modulating the granularity of predictions.