Active Inference & The Emergence of Philosophical Thought
Applications of Karl Friston's Free Energy Principle to Philosophy & Psychotherapy
The diagram atop page 44 of Active Inference - The Free Energy Principle in Mind, Brain, and Behavior by Thomas Parr, Giovanni Pezzulo, and the one-and-only Karl J. Friston is simple but profound.
It shows a brain, a world, and a pair of infinite bi-directional loops that merge to connect the two. In between, a series of statistical inferences, referred to as a Markov blanket, provides the probabilistic distribution of the entirety of human thinking and behavior - cognitive psychology and psychopathology, philosophy, the arts, and political science (not to mention predictive models for the new forms of knowing that are already getting muddled up with human thinking - artificial intelligence and machine learning).
Could it be so simple? No mystical intelligence. No unknown unknowns. Just an elegant theory explaining how sentient beings either alter their internal states or act upon their external world to reduce entropy according to a discreet set of mathematical principles?
Science can be a real fucking buzzkill sometimes.
Regardless, Friston’s work provides insights that not only apply to cognitive processing and traditional psychological approaches to human emotional wellness but also to information processing and decision-making in all message-passing systems of the body. Friston’s assumptions connect the body and mind, providing a unifying principle applicable to both maladaptive worldviews and defense mechanisms (and the cognitive therapies we use to address them clinically) and to the complex system of electrochemical interactions occurring at the synaptic level that underpin our deepest sense of ourselves and the world.
Karl J. Friston’s Free Energy Principle & Active Inference
Karl J. Friston is a British neuroscientist known for developing theoretical and statistical models for cognitive processing in the brain. Most notably, Friston’s work has been leveraged to create predictive models for artificial behavior and learning - artificial intelligence (AI). I was introduced to Friston through the work of South African Neuropsychoanalyst Mark Solms and Canadian Neurophilosopher George Northoff. Both neuropsychoanalysts, Northoff and Solms integrate the fundamentals of the Free Energy Principle and Active Inference into their neurobiological approaches to understanding the brain and within their philosophies of consciousness. Friston’s two central theories, Active Inference and the Free Energy Principle have earned him wide renown, making him one of the most highly cited scientists ever. And as I will discuss later in this article, work is being done to map Friston’s central theses to psychotherapy.
The Free Energy Principle
In short, the Free Energy Principle states that any inferential system (that is - any adaptive system that uses data to draw conclusions or determine action) will operate with the implicit goal of minimizing free energy. Organic systems, from the amoeba to the system of systems that we call homo sapiens, minimize free energy in an attempt to maintain homeostasis. The mammal, for example, experiences an increase in free energy when there is a disruption in the system that regulates the transformation of caloric energy into kinetic energy (it gets hungry) and attempts to minimize this free energy by finding food and eating, thus returning the system to homeostasis.
In the age of AI, adaptive systems can include the inorganic - like The Almighty Algorithm that, for example, scans your retina through your smartphone camera to learn how your pupils dilate, delivering more of what it thinks you want based on what your gaze programs it to know about you.
(In common parlance, of course, The Almighty Algorithm, has entered into the realm of mythology and ideology for many, endowing it with a power beyond its actual capability. The Almighty Algorithm, which claims to process the aggregate of combined human knowledge at lightning speed, suffers not from human doubt, nor the existential dread we experience when we wonder if our conception of ourselves and the world could be faulty. As a result, many interact with The Almighty Algorithm as if it were indeed of superior intellectual power, forgetting that it has been created by humans who, if they do not doubt, suffer from a conceit that, paradoxically, would be embedded in the mathematical and logical inferences with which they create artificial mechanisms to emulate authentic human intelligence.
Fear not The Almighty Algorithm. Those with the intellectual capacity required to develop it, humble scientists as they most certainly are, must also be aware that they have to some order of statistical got it wrong. Users must remember this too.)
I digress.
Active Inference
Friston’s theory of Active Inference posits that inferential systems perform the above minimization of free energy in two ways: 1) updating a perception about the data that disrupts homeostasis in the system (belief updating), and/or 2) taking action upon the system within which it is embedded.
In the language of human psychology, Active Inference encompasses both the cognitive and the behavioral. According to Friston, when faced with a disruption in the homeostasis of our bodily system, we will either update our beliefs regarding the perception of sensory data or act upon our environment to reduce the free energy.
This gets interesting in the case of the hyper-cognitive human animal who likes to convince itself that it can foresee the future. Humans strategize based on their beliefs about the world and how they interact with it (often neurotically or obsessively). According to Active Inference, this attempt to reduce free energy by foretelling the future and creating a behavioral strategy to address it correlates to the system planning to minimize surprise in the closed system and eliminate an unexpected introduction of free energy. When we complexify our beliefs about a situation to include all conceivable potentialities, or when we take pre-emptive action to reduce the probability of a disruption, we are planning ahead to reduce the unexpected introduction of free energy into the system that we call our self.
How the Body Makes Decisions
The complementary principles of Active Inference and Free Energy have obvious applications to cognitive neuroscience and even the psychology of clinical mental health. But Friston’s principles go further to address deeper neurophilosophical questions: If, from the materialist perspective, all human thought and emotion are functions of our neurology, what processes motivate neurons to create the complex networks that correlate with human thought and consciousness? How does the nervous system create psychology (or to take it one step further, philosophical thought) out of the binary chemical decision-making of sense neurons?
According to Active Inference, change in a system, be it belief updating or acting upon the environment, is based on message passing between systems and within systems of systems. This means that the Free Energy Principle applies not just to cognitive and brain processing but to any place in the body where neuronal systems interact and pass messages with each other about the assumed environmental states from which the data has been drawn. At the interface of one system and another (be it limbic areas of the brain to the prefrontal cortex, brain stem areas to the limbic system, the spinal cord to the basal ganglia, or the “leg neuron connecting to the hip neuron”) message passing occurs in the form of probabilistic expressions about the possible introduction of free energy and surprise. In other words, when a system receives sensory data, a prediction of accuracy occurs with a degree of potential error. Message passing occurs between systems as these probabilistic expressions are communicated across various system interfaces upward through a hierarchy of systems - with less complex systems passing messages “up” the hierarchy (Parr et al, 2022, p.83). In their most rudimentary state, our nervous systems are like on-off controllers of reflexive actions (p.84) that operate on upwardly communicating message junctures.
For example, hearing is based on the receptive neurons associated with the movement of tiny hair follicles in the ear canal. When someone speaks, sound waves move the air across the hair, triggering a series of binary responses in these neurons. (For neurons, the binary decision is “Yes, trigger the next neuron in the sequence” or “No, don’t”). The sensory data delivered by the auditory system passes messages upward through the hierarchy of systems, triggering a cascade of activity along neural systems called reflex arcs. These reflex arcs eventually reach the brain, where cognition is believed to occur.
Probabilistic message passing occurs through the system of systems that is the human body. At each interface in the complexity hierarchy, “decisions” are made to either update the system’s predictive coding or act upon the world. When the Active Inference dictates that a behavioral outcome is warranted, the message is passed up the hierarchy to another interface. Downward from the cortex, and upward from motor neurons that pass through the brain stem and basal ganglia, interpretations of probability occur. The cumulative message passing across these systems leads to the sum of human thinking and behavior.
Parr et al (2022) put it this way:
Sensory afferents enter the brain stem or spinal chord and synapse on motor neurons. Descending predictions of the sensory input are propagated from the cortex to the motor neurons, whose output depends on the difference between their cortical and sensory input (p. 83)
How accurately does the predictive messaging from the neurons coincide with what the cortex is accustomed to expecting from this message passing pathway? Depending on the divergence of this data from what is expected, a probabilistic prediction occurs that will trigger either change in the synaptic expectation of the cortex (belief updating), or a behavioral response, an action upon the environment that has the highest probability to maintain homeostasis for the organism.
I will avoid most of the mathematics (for my own sake and yours). But suffice for now to say that Friston leverages Bayeses statistical theorum which assigns probability to observed and unobserved parameters that, when applied to neurology, correlate to the past experience and potential unknown future experience. The boundaries of the totality of potential Active Inference outcomes create a Markov Blanket which refers to all of the variables and statistical possibilities that impact the interactions between the system and the system in which it is embedded (p.43)
In the Brain
As we know, the central nervous system activates the brain as it connects to the spinal cord and basal ganglia. Parr et al explain that the input from the dorsal horn of the spinal cord meets the predictions encoded in the motor cortex. This results in a “proprioceptive prediction error” that drives reflexive muscle activity across the interface between the two neuronal systems (p.90).
Activity in the basal ganglia connects to neuronal networks located physiologically near (subcortical) and far (cortical). Active Inference decisions inhibit or disinhibit neural pathways, activating various regions and processes throughout the brain. Neurotransmitters come into play by balancing the selection of counter inhibitory neural pathways. Parr et al sum up the work of dopamine in the following:
“Dopamine tends to promote the direct pathway and execution of specific policies - presumably those associated with the lowest expected free energy. In contrast, low dopamine might be expected to favor context-sensitive priors in the indirect pathway, whose role is to suppress implausible policies in a given context….Striatal dopamine can be thought of as modulating the balance between inferring what to do and what not to do” (p.94)
Dopamine further contributes to signal confidence by increasing the gain (signal strength) of a prediction and communicating a message of precision. In the case of learning, for example, increased precision correlates with increased neural response in the receiving system and an updated belief. Active Inference’s dual belief updating and acting upon the environment comprise the strength (gain) and precision of message passed between neuronal circuits and reflexive motor action, respectively.
In the cortex, stereotypical circuits of neural connectivity occur in what are called cortical columns that span the cortex from the areas closest to the top of the brain, to the areas closest to the limbic areas of the brain. In the cortex too, message-passing hierarchies exist as decision-making complexifies from the simplicity of habits to the complexity of strategizing goal-oriented explorative and exploitative actions (p.101).
When it comes to our cognitive schemata (beliefs), Parr explains:
“The substrate of these beliefs is the efficacy of synaptic connections between the neural populations representing time-varying variables (like hidden states or outcomes). When we observe an outcome that we believe was generated by a given state, we can update beliefs about the parameter connecting the two, reflecting an increase in the probability of them co-occurring in the future. In other words, we get a strengthening of the synapse between the two populations of neurons” (p.99)
Our beliefs about the world are not discreet neural networks, but systems of subordinate connections between variables influenced and updated over time by the prediction updating and reflexive action of Active Inference computations.
From the neurophilosophical perspective, perhaps beliefs and values correlate to lower-level decision-making in the cortical column hierarchy. Belief systems and biases are almost habitual in that they are generally not updated on-the-daily but settle into explicit memory as a short-hand framework that subtly underpins more complex life decisions that is only brought to conscious awareness when deliberately called upon, either in the course of cultural tribalism or a deliberate religious, spiritual, or political exercise.
Where do Thinkers Come From?
In an earlier article, I discussed how neuropsychoanalyst Mark Solms carries the torch of affective neuroscience and affective consciousness lit by the late great Jaak Panksepp. The affect-primary theories of Solms and Panksepp dovetail nicely with Friston’s conclusions about message-passing across neuronal system interfaces and maintaining homeostasis. Panksepp also saw the brain as a self-equilibrating system and conceptualized the self as a primordial homeostasis-seeking mechanism rooted in the oldest regions of the brain (the basal ganglia and its connection to the spinal cord). Panksepp argued that this mechanism was present in all mammalian brains and was likely also present in more primitive organisms, though without the reflexive awareness associated with consciousness of the self.
However, Mark Solms follows the Active Inference thread of thought to a philosophical impasse in his 2020 book The Hidden Spring (Solms, 2020). After providing an overview of Friston’s work and how it connects to neuropsychoanalysis, Solms identifies that the juncture and alteration of internal belief states and external states as perceived by sensory apparatus, do not explain the emergence of the ultimate mystery of human psychology, the emergency of the self in reflexive self consciousness. The subject of consciousness (be it the self, the soul, or otherwise) has been the focus of thought within numerous philosophical schools, including Buddhism, the Hegelian dialectic, and post-modernism, that demonstrate the paradox of the subject that both objectifies and negates itself.
How does this style of thought fit in with the pragmatic, mechanistic, and mathematical decision-making model of Active Inference? Does the Free Energy Principle reduce the rich philosophical world of human consciousness to nothing more than belief states interacting with sensory that will be either updated or acted upon depending on our organism’s need for neurobiological homeostasis?
Where exactly do philosophical questions come from?
In the chapter on the Free Energy Principle in The Hidden Spring, Solms brings the reader along his line of thought, how he discovered the work of Friston and Carright-Harris, and how his interest in the potential of the free energy principle to reduce physics, neurobiology, and consciousness into a united theory led him to a question similar to the one above: Where does the phenomenological experience of subjective consciousness come from?
Solms explains:
“Biological self-organizing systems must test their models of the world, and if the world does not return the answers they expect, they urgently do something differently or they will die…This is how question-asking arises; self-organization brings participant observers into being. The question that a self-organising system is always asking itself is simply this: ‘Will I survive if I do that?’ The more uncertain the answer, the worse for the system.”
Solms’ use of the expression “participant observers” highlights the paradox of the self in terms of Active Inference at the cognitive level. A participant observer emerges to fill Active Inference’s need for question asking at the individual cognitive level. Questions must be asked to create predictions about free energy.
Philosophy as Active Inference Exaptation
Is this enough to explain the emergence of complex spiritual and philosophical thought? Surely such thinking extends beyond its survival advantage for maintaining organismic homeostasis. Or is it that the philosopher or spiritual thinker represents a glitch in the Active Inference process?
It has often been said that only the privileged have the luxury of philosophizing. Does it represent an exaptation (a process that was selected for survival purposes finds another function in the absence of its original evolutionary purpose) emerging from societal circumstances that reduce the necessity for survival-based Active Inference questioning?
Stay with me.
Take the example of post-traumatic cognition. Traumatic events induce learning at the neurological level - message delivery expectations based on a traumatic occurrence that is subsequently reinforced by a toxic environment. Even when removed from the threatening environment, the subject continues to ask Active Inference questions based on the message signaling of a wounded neurological system. In this case, questions continue to be formulated to reduce expected free energy based on the expectation of a threat that no longer exists in the environment. This contributes to the sort of compulsive thought (overthinking, anxiety, intrusive ideation) and action (substance abuse, maladaptive defense taking) that we associate with post-traumatic thought - cognitive syndromes that no longer serve the maintenance of organismic homeostasis for the individual. In short, there is a glitch in the planning for surprise based on the faulty belief that the threat could at any moment introduce free energy into the system of the mind.
How does this relate to philosophical thought?
The thought upon thought associated with philosophy could be conceptualized as an exaptation of Active Inference questioning. It serves no survival purpose. Although the philosopher no longer needs to plan for any disruption in the homeostasis of the mind, questions continue to be asked (like the continuation of traumatic styles of thinking that linger after the threat no longer exists) as if the mind is planning for an introduction of free energy that will never occur.
If the aspirational attempt to create a unified theory of truth, knowledge, morality, etc (as in the case of the “provisional aspiration”) is an attempt at greater and greater entropy of belief, it could be seen as the mind’s attempt to reduce the surprise free energy that results from encountering a situation that one is philosophically unprepared for (as in the case of existential self-doubt or embarrassment). Hypothetically, if one’s philosophical stance is robust enough, no human experience would fall outside it.
While the philosopher in optimal circumstances creates a work of literary art or cognitive usefulness (community wisdom), a similar inquisitive day-dreaminess could be interrupted by a necessity to act to maintain homeostasis over belief updating, thus distracting the philosopher from their belief updating work - the creativity that is philosophical thought, the inspiration, the dream before enactment.
Active Inference in the Therapy Room
Fristonian thought equates psychopathology with a “failure of belief updating” (Parr et al, 2022). Psychiatric disorders are seen as computational disorders of decision-making in the body. In Active Inference - The Free Energy Principle in Mind, Body, and Behavior Parr et al link psychopathology to “aberrant prior beliefs” and provide a table that describes how the Active Inference process, based on faulty prior beliefs, can lead to the behaviors that we associate with mental disorder (Parr et al, 2022, p.186).
One example is addiction, research by FitzGerals and Schwartenback (Schwartenbeck et al, 2015) that connects addictive behaviors with prior beliefs about the likelihood of receiving a promised reward. Low confidence levels in the expectation of reward result in obsessive and compulsive relationships with the substance or activity. Adjusting prior beliefs about the availability of reward could reduce the desperate nature of the hedonistic grasp, such that recovery looks behaviorally more like moderated use of addictive activities and substances, rather than the neuroticism of rigid abstinence or obsessive sublimation.
The table also lists psychopathy, connecting belief models about interpersonal behavior with personality disorder. Active Inference may work upon faulty prior beliefs about how individuals generally respond to interpersonal behavior. In the studies cited, researchers mapped belief about “the degree to which beliefs about self-worth depend on decisions to be charitable versus selfish and sensitivity to the approval of others” (Parr et al, 2022) to the emergence of the behavior traits commonly associated with psychopathy.
So, how do maladaptive false beliefs come about? The Active Inference model provides a fascinating theory.
In the message-passing computations mentioned above, the data in each message passed appears to be encoded with a measure of precision - a likelihood that a correct inference about the system’s environment can be drawn from the sensory data. When it comes to cognitive schemata, the precision of a belief is potentiated by neurochemicals. Parr et al suggest that the cholinergic system (part of the nervous system that responds to the neurotransmitter acetylcholine) and the serotonergic system (responding to the neurotransmitter serotonin) communicate precision to sensory data sets. These neurotransmitter systems could potentiate a message such that it is received as having a higher precision than the sensory data suggests. Parr et al refer to this as “overinterpretation of (potentially noisy) sensory input” (page 189).
The above aligns with clinical methods that facilitate recovery from interpersonal trauma. In my previous article about Allan Schore’s Right Brain Psychotherapy, I discussed modern trends in neuropsychotherapy that highlight the importance of recreating the neurochemical situation of the dynamic mother and child attachment dyad in the therapy room. When the positive social messages of therapy are potentiated with oxytocin (mimicking the neurochemical environment of the early social development of the brain), they become more fixed in the client’s brain, in a sense chemically overpowering prior beliefs created by early trauma with an “overinterpretation” of the sensory data associated with the prosociality of the therapy room.
In the therapeutic chamber, the counselor is the administrator of free energy by alternately inserting and reducing surprise (in the traditional social sense) across the social synapse. In the artificial environment of the therapy room, they act differently than what the client expects to expect from normative social interaction. The surprise is registered by the client’s sensory apparatus - sight, auditory, bodily, processing at often unconscious levels- disrupting and facilitating emotional and neurochemical rupture and repair (Haskayne et al, 2014). This is the art of creating a “safe but not too safe” therapeutic environment (Solomon & Siegel, 2017). The client’s ability to respond to ever-increasing levels of entropy (free energy) with either spontaneous belief updating or assertive action upon the environment (assertive communication, boundary setting, self-empowerment) corresponds to mental wellness.
In the article “Surprise” and the Bayesian Brain: Implications for Psychotherapy Theory and Practice (2019) Jeremy Holmes and Tobias Nolte state that “the brain abhors surprise”. And prescribe a “bio-behavioral synchrony” corresponding to Allan Schore’s intersubjective field that attempts to recreate the mother-child dyad using right brain dominance in both client and therapist to moderate intense affect.
Integrating Active Inference Holmes and Nolte say this about the client:
“They are in the grip of perceptual distortion and ingrained prediction errors, driven by the need for a modicum of attachment security, howevery dysfunctional…In the presence of a modulating, moderating, affect-buffering therapist, surprise/energy unbound becomes tolerable and, when therapeutically scrutinized, extends the repertoire and range of a person’s counterfactual realities, i.e., priors…Reducing prediction error is a complex multi-level and recursive process that reverberates up and down a series of interconnected message-passing hierarchies” (Holmes & Nolte, 2019)
Friston has received a lot of praise for his theory of Active Inference. Rightfully so. Active Inference and the Free Energy Principle provide a simple and elegant unifying principle for all systemic decision-making. The mind runs wild with potential correlations and applications.
I’ll leave you with one question for the rest of your day - an introduction of free energy to disrupt your homeostasis and rev your Active Inference engine. That thing you’ve been thinking about. You know the one. That task. That goal. That obsession. That fear. That girl. That boy. That dream. Will you update your belief, or will you change your world?
Have fun.
References:
Haskayne, D., Larkin, M., & Hirschfeld, R. (2014). What are the experiences of therapeutic rupture and repair for clients and therapists within long‐term psychodynamic therapy?. British Journal of Psychotherapy, 30(1), 68-86.
Holmes, J., & Nolte, T. (2019). “Surprise” and the Bayesian brain: implications for psychotherapy theory and practice. Frontiers in Psychology, 10, 592.
Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active inference: the free energy principle in mind, brain, and behavior. MIT
Schwartenbeck, P., FitzGerald, T. H., Mathys, C., Dolan, R., Wurst, F., Kronbichler, M., & Friston, K. (2015). Optimal inference with suboptimal models: addiction and active Bayesian inference. Medical hypotheses, 84(2), 109-117.
Solms, M. (2021). The hidden spring: A journey to the source of consciousness. Profile books.
Solomon, M. F., & Siegel, D. J. (2017). How People Change: Relationships and Neuroplasticity in Psychotherapy (Norton Series on Interpersonal Neurobiology). WW Norton & Company.