Mark Solms. ‘Feelings are real, and we know about them because they permeate our consciousness. […] From their origin in some of the most ancient strata of the brain, they irrigate the dead soil of unconscious representations and bring them to mental life.’ (Solms, 2021).
1. Phenomenal experiences of feelings, which provide meaning.
2. Phenomenal experiences are qualitative - not quantitive.
3. Phenomenal experiences arise from a hidden spring, the subcortex (e.g., brainstem areas).
Cleeremans and Tallon-Baudry: only because ‘conscious agents experience things and care about those experiences’ they are ‘motivated to act in certain ways’ and ‘prefer some states of affairs vs. others’ (Cleeremans and Tallon-Baudry, 2024).
Anil Seth. Phenomenal experiences emerge from embodied, control-oriented predictive regulation and that ‘experiences of having a body, and of simply “being” a body, the many and varied elements of selfhood are Bayesian best guesses, designed by evolution to keep you alive’ (Seth, 2021).

Marc Solms provides an interesting take on consciousness, developed together witch Karl Friston. He says /read quote/
- (1) Phenomenal experience result from feelings that have intrinsic meaning for my Self in the form of positive or negative valence, which makes them subjective.
- (2) Phenomenal experiences are not just processes that could be modeled as quantitative regulation processes using continuous variables (like reflex arcs or homeostatic regulation). Instead, they are qualitative states that require categorical distributions.
- (3) Pheonomenal experiences do not arise from cortical but from subcortical regions. It is the The Hidden Spring of consciousness. This is evidenced by several lesion studies in clinical cases.
- I really recommend this book because it provides an easy access route to active inference, and also because Solms is not so overly Freudian as in other books.
Dobzhansky famously proposed: Nothing in Biology Makes Sense Except in the Light of Evolution - in evolutionary and cognitive neuroscience we also aim to justify biological or cognitive phenotypes with evolutionary, ecological arguments. The hardware that could provide phenomenal experience is metabolically expensive. So we need a good reason for it. Keeping us alive in a complex ecological niche, far away from thermodynamic equilibrium is usually one of the best reasons.
Anil Seth argues similarly and provides computational details on interoception using Bayesian generative models, predictive processing, and active inference. A framework that I will use as well here.

Convenient Ecological

& Computational Theory

referring to Thomas Metzinger (2024) The Elephant and the Blind

referring to Brooks, R.A., 1990. Elephants don’t play chess. Robotics and Autonomous Systems 6, 3–15.

The elephant breaks self-serving models of consciousness.
The elephant breaks habitual generative models.
- Usually, generative models are hierarchically high, laterally wide, and temporally deep.
- What kind of model architecture is needed for MPE?
The elephant breaks active inference.
- Usually, generative models are for action-oriented cognition for adaptive Bayesian self regulation and epistemic Bayesian inference.
- What is the alternative to these kinds of models during MPE?
The elephant breaks the free energy principle.
- Goal: Optimal Bayesian models and energy-efficient neuro-computational message passing $\leftarrow$ Surprise minimization and complexity reduction $\leftarrow$ Minimization of (variational and expected) Free Energy
- If any, what kind of free energy minimization is done during MPE?

Computationally, MPEs do not work with habitual, business as usual, generative models. This is what we will discuss first. Do they break the model architecture, are they a specific singularity, an edge case - or are they compatible with traditional generative models.
What is the alternative to the active inference proposal that cognition is action-oriented for adaptive self regulation? If there is no action or self. Computationally, we usually talk about a generative self model. Can there be modeling without a model? There needs to be some kind of additional model outside the self-model. I will argue that the brain can be partitioned in a way that would allow for this.
When exploring the computational requirements we will reach a point where it seems that the models violate the generalizability of the Free Energy Principle. Can there be cognition, or life as we know it, where Free Energy is not minimized. I will argue that even during MPE, the brain minimizes Free Energy - by reducing model complexity at the highest abstraction and outside the self model.
So you see there is a lot at stake for me - I would need to retract papers or - even worse! - update my teaching slides.

	Habitual Cognition	MPE Cognition
Lateral Width	High - multiple (counter factual) hypotheses in parallel.	Low - for focussed attention (Easy, 4.1)
Temporal Depth	High - to predict the consequences of potential actions.	Low - for staying in the moment (Easy, 4.2)
Hierarchical Height	Adequate - to model the data (high accuracy, low complexity)	Too complex? (Harder, 4.3)
Self-modeling	Yes.	No. (Oh dear, 4.4)

for details see Sladky, R. (2024). Of Hidden Springs and Endless Oceans. 10.31234/osf.io/9tj3s

From bottom-up processing to top-down predictions. e.g., An active inference perspective for the amygdala complex (Sladky, R., 2024, TiCS)
- The amygdala links interoceptive regulation (via inputs from brainstem and interoceptive models in the insular cortex) to exteroception (via temporal lobe and thalamus)
- Amygdala output is crucial for modifying behavior via brainstem (e.g., fast, stereotypical threat responses and regulation of the autonomous nervous system), ventral striatum (e.g., biasing action selection), prefrontal cortex (e.g., active avoidance requiring long-term planning) and hypothalamus (e.g., somatic arousal via hormonal cascades).
Subcortical cognition. Paradigm shift in neuroscience that subcortical (limbic) areas are on top of the computational hierarchy (e.g., Chanes, L. & Barrett, L.F., 2016. TiCS)
Brain evolution and development. Finding a link cortical development to neuroimaging (Bernhardt et al., 2022), such as the dual-origin hypothesis (Sanides, 1962).

My neuroscience research focusses on the amygdala.
I have recently reviewed the amygdala’s role in creating predictive models for homeo-/allostatic control and perception from an active inference perspective (Sladky et al., 2024) - which appears to be well in line with Mark Solms proposal.
Read
The subcortex plays an important role: The amygdala is a central hub for multimodal/multisensory integration. It is on top of a hierarchical active inference model. This explains why amygdala activation (e.g., prediction errors from deviations from homeostatic preference priors) leads to strong changes in behavior (e.g., avoidance or approach behavior) and subjective experience (e.g., fear and other emotions).
Placing subcortical areas on top of the hierarchy and not the prefrontal cortex is an important paradigmatic shift. The prefrontal cortex allows for functions that have traditionally been called "higher" cognitive functions, such as language production, based on the notion that they are hallmarks of "higher" animals like humans. We can explore this more in the discussion but remember Rodney Brooks paper Elephants don't play chess. Chess is probably not be the most relevant brain function. Making good allostatic predictions is.
Finally, understanding evolutionary trajectories of brain development have gained interest in neuroimaging, es. Among them the dual origin hypothesis of cortical development suggests that two distinct brain networks emerged from different expansion points (Sanides, 1962), potentially leading to marked functional differences. One system originated from the olfactory system and an amygdala-centered expansion gradient towards ventral cortex, which could enable interoceptive self-modeling for habitual interactions with the body and the environment. The other system is hippocampus-centered, expanded towards dorsal cortex, and could enable allocentric forms of cognition and experience. Let's look at it these two systems.

1. From the amygdala to the sensorimotor cortex via the insula cortex and operculum. This gradient would be a good candidate for interoceptive functions and self-regulations. The expansion gradient terminates is the ventral part of sensorimotor cortex, which are relevant for face movements and sensations, important parts of affective behavior and experience. 2. From the amygdala to the posterior temporal lobe. This gradient reflects the ventral perceptual stream. The so-called What?-pathway in vision is for identification of visual stimuli from simple concrete (posterior part) towards more complex features (anterior part), and relevance detection (amygdala). Importantly, it exhibits regions that are optimized for specialized feature detection (e.g., faces). This expansion gradient plays and important role in exteroception. We assume the basolateral amygdala plays a role in tuning hierarchical feature selection (Sladky et al., 2024). A similar gradient exists for auditory processing. 3. From the amygdala to the ventral prefrontal cortex (vmPFC, vlPFC), which we know play a role in what is traditionally called prefrontal amygdala regulation (Sladky et al., 2021a, 2015).

1. From hippocampus to anterior cingulate cortex and dorsal prefrontal areas (e.g., dorsolateral prefrontal cortex). Functionally, this gradient plays an important role in the formation of declarative, associative memories, and relating new to old information (Menon, 2016). After acquisition, as soon as memories are established in cortical areas, the functional relevance of the hippocampus is decreased (Eichenbaum, 2017) and, eventually, no longer required during the retrieval processes (Qin et al., 2014).

2. From hippocampus to parahippocampal cortex and parietal lobe. Historically, hippocampus has been studied extensively in the context of navigation. Place cells in the hippocampus (and grid cell in the adjacent entorhinal cortex) dynamically encode the current location. During memory consolidation, parietal cortex encoding changes to a more permanent and action-oriented format (Whitlock et al., 2008). Following an abstraction gradient, hippocampus encodes abstract coordinate information, while parietal cortex is crucial for more concrete, categorical spatial relationships (Baumann and Mattingley, 2014).

3. From hippocampus to medial cingulate cortex and dorsal sensorimotor network. The sensorimotor network is a mostly cortical network that includes primary and sensory-motor areas as well supplementary motor areas (Yeo et al., 2011). As such, the world model from the hippocampus network enables coordinated, goal-directed motor actions.

Giaccio, R.G., 2006. The dual origin hypothesis: An evolutionary brain-behavior framework for analyzing psychiatric disorders. Neuroscience & Biobehavioral Reviews 30, 526–550.

Common in both systems is an increase in neuroanatomical complexity from their expansion points. They reflect a progression from older structures with fewer neural layers towards newer structures with more differentiated layers at the end of the expansion gradient: from the primordial limbic zones referred to as allocortex (amygdala and hippocampus) over three-layered paralimibic areas (piriform, entorhinal, parahippocampal, and the ventral cingulate cortex), to areas without (agranular insular and pre-central cortex) or only rudimentary granular layer IV (dsygranular insular, dorsal cingulate, and orbitofrontal cortex), towards the newest six-layered cortex, known as isocortex or neocortex, with the highest laminar differentiation only found in primates but not in rodents.
Also check out Karl Fristons 2024 Cerebral Cortex paper: From active affordance to active inference: vertical integration of cognition in the cerebral cortex through dual subcortical control systems

For richness and flexibility in perception and action you need lateral width. E.g., homeostasis (first-order cybernetic control loop) vs. allostasis (more degrees of freedom).
The brain is a phantastic organ. Intrinsic connectivity and dynamics dominate in the brain.
- Focused attention. Tuning the generative model to specific (task) requirements (e.g., detecting a face in the crowd).
- Inattention. Increased likelihood that other models dominate, a phenomenon known as mind wandering.
Neurocomputational hypothesis. Different models are iteratively selected and optimized to avoid being stuck to a model where no optimization progress can be made.
- Boredom or frustration. Not making progress could bring previously downregulated interoceptive/exteroceptive models into the foreground for evaluation.
- Mind abhors a vacuum. Explaining away all prediction errors appears to be an unfavored state. Given the system’s intrinsic goal for model optimization there is always the tendency that this vacuum is filled by remembering episodic memories and creating anticipatory plans on what we need to do instead of meditating.
- So, the phenomenal feeling of unconstrained openness in MPE arises from focused attention and not from open, unconstrained cognition in the form of mind wandering.

Many meditation techniques require focused attention on interoceptive perception like breathing. Suppression of alternative models makes the dominant model an optimal interoceptive model of respiration. In parallel, attenuation of prediction error in interoceptive regions such as the insula and, possibly, the attenuation of phenomenal self-experience.
The insula does not create a static mapping of an interoceptive self-model. Instead, insular activation is highly dynamic in response to interoceptive data and task demands.
Focused attention should prevent exploratory top-down selection of alternative models, which have not yet been attenuated. Low prediction error in the insular cortex leads to low prediction errors in the central amygdala $\to$ decreased action initiation and basolateral amygdala predictions along the exteroceptive parts of System-A.
- An experimental intervention could be the downregulation of the basolateral amygdala using fMRI neurofeedback training (Watve et al., 2024) or non-invasive brain stimulation.
Next challenge. Agentive attentional top-down effort could lead to the strong experience of agentive control and self-identification - Instead, the goal would be lateral inhibition of alternative models and not top-down control.

I argue that MPE requires both focused attention and non-agency. Focused attention (i.e., ‘voluntary focusing attention on a chosen object in a sustained fashion’), vs. open monitoring (‘involves non-reactively monitoring the content of experience from moment to moment’) (see Lutz et al., 2008 for a neuroscientific review) but non-agency is a problem for active inference (see Dark Room Problem).

	Short temporal depth	Long temporal depth
Exteroception
Epistemic	Perception as hypothesis testing	Exploration as testing counterfactual hypotheses by actively sampling the world
Instrumental	Regulation by taking actions to ensure the world confirms to expectations	Anticipation by simulating the consequences of future actions
Interoception
Epistemic	Perception as instrumental model confirmation by directing attention to predicted interoceptive sensations	Exploration, e.g., self-evidencing behavior (Perrykkad and Hohwy, 2020) such as self-touching or altering respiration
Instrumental	Regulation by habitual behavior	Anticipation by goal-directed behavior

What happens during meditation aiming for non-action, non-goal directedness, and presence in the moment?
- Shrinking temporal depth of the model. Being in the moment entails reduced planning of future actions. I consider this a necessary first step, but you also need to prevent agentive exploration (like opening the eyes) or regulation (like scratching). Attenuation of prediction error in the interoceptive and exteroceptive model (high accuracy) could reduce incentives for actions.
- This highly accurate self-model would be optimal only for a short time. It does not feel right and increases need for self-evidencing actions, i.e., attention shift (focusing on breathing) or behavior (touching oneself). Absence of prediction errors also drives explorative behavior to minimize expected free energy using habitual or goal-directed actions mediated by striatal areas.
- Downregulating the striatum could reduce goal-directed and habitual epistemic and instrumental actions. Focused ultrasound of the caudate nucleus has been used to support meditation practice (Cain et al., 2024).
MPE is accompanied by a loss of time perception (Gamma and Metzinger, 2021), which could be caused by attenuated interoceptive signals that provide temporal structure in System-A. Additionally, hippocampal time coding is thought to be actively constructed as sequences (e.g., of episodic memories).
Next challenge. Attenuating the interoceptive and agentive Self entails that certain higher levels in the hierarchy do not process any meaningful data. These models are not optimal anymore, they are unnecessarily complex. This violates the Free Energy Principle.

For MPE, I propose a model architecture that is low in width (i.e., effortless focused attention prevents mind wandering) and short temporal depth (i.e., an attenuated, well-regulated self-model without anticipation of future actions). However, such a model is unable to increase accuracy.
- The only possible model optimization entails complexity reduction. However, MPE would lead to needlessly overcomplex models as some hierarchical levels essentially encode no relevant data. Do MPE break model optimization goals and present a case beyond the scope of active inference and its free energy minimization imperative?
I argue that MPE might be the phenomenally most simple experience – yet, it requires a computational model of high complexity. Adding hierarchical levels on top of simple regulation loops adds layers of progressively simplified abstraction.
- If I would directly experience only my bodily regulation efforts, e.g., how my nervous cells anticipate the next bolus of oxygen-rich blood, I would assume this experience to be very unpleasant and stressful, maybe comparable to the experience of a panic attack.
- Adding another layer that explains the prediction error on the lower level away removes this stressor and properly calibrates the signal-to-noise ratio to relevant deviations from homeostatic set points, e.g., if the actual deviation of blood oxygenation is larger than expected.
- How many layers can you add to explain away the Self?
- Until you reach a flat line?

Computuational model says: No.

Overcomplex models should not exist for a long time. Phenomenologically, I suggest that this added complexity leads to the experience of calmness, turning into boredom after some time. Based on the Free Energy Principle, we would expect that not being able to minimize free energy would be an unpleasant state (Joffily and Coricelli, 2013).
If life as we know it entails free energy minimization (Friston, 2013), MPE states would computationally amount to being dead. If we dismiss this option, I suggest MPE entails free energy minimization outside the scope of the completely attenuated self-model. But how?
Bayesian model reduction. If meditation has initiated too many levels of computation (e.g., meta awareness) these layers will no longer contain any relevant data to explain the observed variance in sensory data (i.e., a flat line). It should be removed for parsimony.
- Psychologically, this process has been called fact free learning (Friston et al., 2017) and could be related to the Vipassanā (insight) meditation practice (Laukkonen and Slagter, 2021).
Next Challenge. Removing superfluous top-levels from System-A would stop the attenuation of the interoceptive self-model and reinstate the experience of being a body. Likewise, reducing abstract System-H top layers reinstates more concrete episodic memories, mediated by activity of an unattenuated default mode network. These autobiographical memories would likely reinstate their affective self-reference via a reengaged interoceptive self-model in System-A and, consequently, increase temporal depth again as one starts to simulate actions for counterfactual futures.

Alternatively, model complexity is reduced by simplifying the encoding of the modeled distribution – not by reducing the top layer. Phenomenologically, this could lead to the experience of knowledge increase in the most abstract sense.
- In System-A? Speculatively, I assume that this kind of optimization cannot occur in System-A as the flexibility of the self-model is constrained by physiological needs that limit System-A’s degrees of freedom.
- In System-H? System-H is less constrained; its architecture resembles an ideal system for unlimited associative learning, which has been proposed as another evolutionary driver for consciousness (Jablonka and Ginsburg, 2022).
Potential neuronal mechanisms are currently discovered. Predictive models are created in the hippocampus (Stachenfeld et al., 2017) and entorhinal cortex (Ouchi and Fujisawa, 2024) and hippocampus learns to generalize by reducing complexity (Courellis et al., 2024).
I can only speculate about the phenomenal experience of this neurocomputational process but would assume a non-egocentric and non-agentic feeling of open-ended non-convergence and positive valence.
Next challenge. What about the self?

Ego-centricity and ownership is the final challenge. The generative model is perspectival: it describes the computation that needs to be implemented for the agent’s self-regulation. Likewise, the phenomenal model exists for someone (the agent) to do something (self-regulation).
We need the self-model to be stable in focus (low hierarchical width), attenuated and well-regulated (low temporal depth), and at least temporarily over-complex (high hierarchical depth). If free energy minimization is still an imperative the model needs to increase accuracy and decrease complexity somewhere else, i.e., outside its self-model.
Computationally, I suggest the existence of a neurocomputational model outside or parallel to the egocentric self-model, where free energy minimization happens during MPE. I propose that the hippocampus and connected brain areas (System-H) would be a candidate for these processes.
However, for MPE there should not be an involvement of System-H’s lower hierarchical layers, such as the default mode network, which could provide autobiographical details that reactive the self-model.
This peripheral System-H deactivations would be in line with empirical EEG results in meditators who experienced a cessation of conscious experience during meditation where alpha-suppression was strongest over the occipital and parietal regions of the brain (Chowdhury et al., 2023).

Habitual phenomenal experience vs. MPE

We have a hidden spring and an endless ocean.

We are Will and Representation (Schopenhauer, 1818)

We are Will and Wonder (Tool, Pneuma, 2018)

return ~~True~~ something useful

In sum, MPE could require an attenuated System-A but active System-H engaged on its highest, most abstract computational levels leading to the experience of non-self, timelessness, and non-concrete phenomenal content. Explaining away the bodily self-model could even feel ‘supernatural’ (Hohwy and Paton, 2010). Crucially, if System-H were not involved it would not be clear what kind of cognition and free energy minimization takes place to be experienced on a phenomenal level. MPE could be the experience of abstract hippocampal model optimization processes that happen usually without our conscious awareness, which is usually dominated by System-A’s activity in creating an interoceptive model that provides valence for an exteroceptive model and guides actions.
My proposal suggests that we should consider the involvement of two different brain networks during habitual phenomenal and minimal phenomenal experience. What goes on in System-H, this model optimization, is usually to subtle for us to notice. We need to fully attenuate the interoceptive and agentive model in System-A first.

Minimal Phenomenal Experience

Minimal Phenomenal Experience

Of Hidden Springs and Endless Oceans