Could the Free Energy Principle provide insight into philosophy of mind?

“We sail within a vast sphere, ever drifting in uncertainty, driven from end to end. When we think to attach ourselves to any point and to fasten to it, it wavers and leaves us; and if we follow it, it eludes our grasp, slips past us, and vanishes for ever. Nothing stays for us.”

— Pascal’s Pensees: The Misery of Man Without God, French mathematician Blaise Pascal

As Pascal described in Pascal’s Pensees: The Misery of Man Without God, we can make sense of the uncertainty of the world to some degree, but never quite reach complete certainty. When our minds try to guess what may happen next, however, whether we’re determining the outcome of a coin toss or the chance it’s going to rain tomorrow, many times, we rely on forms of reasoning that use different types of inference. Of the theories scientists and philosophers have put forward for explaining how the brain works, Bayesian inference remains popular in addressing these types of uncertainties. This post will explore the possibility of whether the “free energy principle” (FEP), an emerging mathematical formalization describing how the brain makes these decisions, could provide insight into the philosophy of mind, its epistemic implications, and if it could even serve as a grand unified theory (GUT) according to philosophical notions of unity. 

For decades, scientists and philosophers have debated the idea that the brain functions according to Bayes’ theorem. By describing how likely something is to happen in the future based on what you know about the past, the brain can create models of the world. By processing information about the world and creating hypotheses of the future, the brain generates and updates a mental model of the environment. Think of it like having a model of the world, but in your head. You learn about the world and update and refine your models with what you learn by assigning probabilities to hypotheses and updating them as you receive new information. Bayesian inference has been applied to the study of perception, learning, memory, reasoning, language, decision-making, and many other concepts in cognitive science. 

Bayesian reasoning has found its use across areas of psychology. In studies associated with Pavlovian conditioning, our brains use the mutual information exchanged between the data given by our senses in updating our internal models of the world. For cognition, motor control, attention, and working memory, we can make Bayesian predictions of the future. Think of predicting the future like the brain performing simulations on hardware, creating its own world of what’s to come. Can we accurately simulate the future? Only if our model of the future is good enough. 

With this Bayesian idea of the brain as an “inference engine,” the brain acts as a self-organizing system at equilibrium with its environment that can actively change and update its model of the world (through a process called “active inference”) into its next state (the expected state). In the search for grand unifying theories of the brain, the “free energy principle” (FEP), a newly developed formal statement in mathematics, has emerged. This theory states that a self-organizing system (one that changes its own internal structure in response to the circumstances outside of it) that is at equilibrium (or the point of balance between these internal and external states) with its environment must minimize the amount of free energy that it has. In brains, for example, self-organizing behavior has been observed in the nature of the subsystems that make up its connectivity. 

In the context of Bayesian inference, we can evaluate free energy as a function of two things to which the agent of a system has access: the sensory states and a recognition density encoded by its internal states (such as neuronal activity and connection strengths between neurons in the brain). Using these internal models that are used to make predictions with perception, action, and learning, these theories of predictive processing (PP), in which the brain generates and updates a model of the mind, and the FEP have gained significant interest and traction among researchers within philosophy of mind and cognitive science. 

In a similar vein of whether the FEP could unify theoretical work on the brain, researchers wonder whether the FEP could serve as a theory of mind. There has, indeed, been work describing the implications of the FEP which extends PP with a principle of adaptive self-organization to areas of epistemology and the philosophy of mind. There’s also been debate and discussion over the representational or non-representational nature of the FEP, with different proposals that the FEP supports internalist, action-oriented, or enactivist and non-representationalist claims. 

What sort of philosophical interpretation of the FEP may be necessary to account for these claims? Neuroscientist Karl Friston, the main proponent of the FEP, has argued how, when the principle is applied to neural functioning, it leads to the generation of efficient internal representational schemes and reveals the deeper rationale behind links between the features previously mentioned (perception, inference, memory, attention, and action). As the principle functions on different hierarchical levels throughout the brain (from the fundamentals of individual neurons up to the way different parts or regions of the brain interact with each other), the morphology, action tendencies, and neural architecture that result from the system itself can all be seen as expressions of the FEP. 

According to the work of Giovanni Pezzulo, researcher at the National Research Council of Italy, and Matthew Sims, graduate student in philosophy at the University of Edinburgh, we can distinguish between four different types of representational schemes that FEP may support: focusing on the internal aspects of a model with respect to organization (i.e., separating a variable within the system from what it represents outside the system), structure (i.e., having a representation structurally similar to the state of the world it stands for), content (i.e., having internal models that encode the environmental or sensorimotor contingencies, as in, how they describe the relations between actions and the changes in sensory signals that make up sensory experience and awareness) and function (i.e., how they’re used before or in the absence of external events), illustrating the FEP’s potential significance to the philosophy of mind. 

On top of this, the authors bring up the question of the FEP’s relevance to whether organisms have (or need) internal representations, a heated topic in the context of PP. Does the FEP entail internal representation? They arrive at the conclusion that, regardless of the answer, the FEP has been (or can be) used to create models that fulfill the requirements of some kinds of representation. This can reveal assumptions on the notion of representation used in creating some ground in solving the debate on whether organisms need internal representation. 

However, such philosophical interpretations of the FEP still need development. Professor Javier Sánchez-Cañizares of the Mind-Brain Group, Institute for Culture and Society, University of Navarra, has written on how, as a theory of mind, the FEP uses a dual information geometry that allegedly explains epistemic access to the world based on prior dual assumptions. As the FEP relies on non-fundamental, relative system-environment separation, a “double circularity” arises that proves harmless for the FEP as it remains a principle limited to science, giving epistemic advancement to life sciences. Philosophical interpretations of the FEP, such as Markovian Monism, cannot use these types of circularities in their reasoning. Should these philosophical assumptions remain open to criticism that requires justification, Sánchez-Cañizares argues, the science can rest soundly upon them. 

Should it serve as a GUT, the characteristics of the FEP could be compared to those of unifying theories in other areas. Similar features underlying unified theories in different areas of work could point to the idea it could be a GUT.  For example, when you take a look at neural Darwinism in biology and information theory and optimal control theory in physics, what do they have in common? Optimization. By comparing the themes and principles underlying this sort of optimization in other areas and why they’re promising, Friston has argued how, when optimizing certain quantities (value, expected reward, expected utility) as they correspond with surprise (prediction error, expected cost), the same quantity that is optimized keeps emerging. This holds true under the FEP as well. 

Parallels between Bayesian inference and evolutionary dynamics, for example, could imply that there are evolutionary processes that could be rethought of as algorithms for Bayesian inference. The working memory can be modelled as a particle filter, an algorithm for estimating probabilities through a type of approximation updated over time. It figures out which hypothesis “fits” the probability of the observed data the same way evolution selects for individuals most “fit” to survive and produce fertile offspring. If I arrived at a restaurant and realized I forgot my wallet, which is more likely? That it’s in my car or in my home? The “fitness” of each hypothesis is how likely the data fits. Optimizing the expected reward, we update our given evolutionary model appropriately. 

Though we only use working memory for a few objects or events at a time, it’s correlated with general intelligence, giving it a role in higher-order cognition. We use it for forming representations in our minds, and, in evolutionary terms, we can prefer some information over others in a way similar to evolutionary processes selecting the best individuals. When memories degrade, they drift due to time or interference, becoming less and less accurate. Memories can also mutate, letting them sway from the truth. By comparing Bayesian inference to these concepts from evolution, we see that Bayesian inference could’ve evolved for similar purposes – determining the best models to fit our environment. We need a representation of the world to survive in it. 

Creativity could’ve come about the same way. There’s a “zoological evolution” of nature and the evolution of ideas when we perform creative activities, psychologist William James wrote in his 1880 lecture “‘Great Men and Their Environment.” In it, he writes, “new conceptions, emotions, and active tendencies which evolve are originally produced in the shape of random images, fancies, accidental out-births of spontaneous variation in the functional activity of the excessively unstable human brain, which the outer environment simply confirms or refutes, adopts or rejects, preserves or destroys – selects, in short, just as it selects morphological and social variations due to molecular accidents of an analogous sort.” 

Writing a story means synthesizing models of the world and optimizing them. Creativity, like evolution, can be thought of as an “evolution of thoughts.” Take a look at something from all perspectives. A poet may ask, “how might a raven be like a writing desk?” Go through the possibilities of how writing desks can relate to ravens. Notice any patterns, similarities, or associations? How would they relate to each other? Create a model, take note of what you see, and optimize your model. 

With these similarities between Bayesian inference and evolution, there could be a deep connection between both processes. The way one thinks may depend more upon ideas developed in ways borrowed from both evolution and Bayesian inference. How closely related are the two of them? As humans start coming together in the post-pandemic era, can we make sense of the world the same way? As our brains evolved to form predictions together, as though we were a collective, the concepts from evolutionary theory and Bayesian inference could come together as well. Such a unified theory could then revolve around an epistemological notion of unity with similar goals of explanation as a GUT.  

Hurdles to using the FEP as a GUT exist, though. In particular, the FEP faces issues in an epistemic context. Matteo Colombo, Associate Professor in the Tilburg Center for Logic, Ethics, and Philosophy of Science, and in the Department of Philosophy at Tilburg University, and Cory Wright, Professor of Philosophy at California State University, Long Beach, have argued that both the epistemic status of the FEP and its precise role in biological and neuroscientific theorizing is “opaque.” Among its issues are the different formalisms and formulations, the changing ranges over which it has been applied in different contexts, the ways in which the FEP relies on undefined terms and stipulative definitions, and the lack of clarity in the logical structure of reasoning involved in the FEP. In their critique of the FEP as a GUT, they cite the example of the activity of mesocorticolimbic dopaminergic (DA) systems in which the different explanatory models of DA systems justify explanatory pluralism and show that a single overarching GUT doesn’t follow from this scientific progress. 

They argue that, as a type of error correction with Bayesian inference, the FEP, in its conceptual simplicity and powerful formalism, provides a rationale for the Bayesian cycles of perception and action that can be used in unification and scientific reduction of other explanatory principles such as those used in psychology. This framework, ever-present in cognitive science, can be described as the relationship between how feed-forward connections convey information on the difference between what was expected and what was obtained (i.e., prediction error) between an agent and its environment while feedback connections provide the predictions from higher processing stages to suppress prediction errors at lower levels. 

In the context of complex systems, those we often call “wholes greater than the sum of their parts,” the FEP’s explanatory power is still quite limited. Yet, understanding the role it may play in related questions or topics of what might constitute consciousness or self-awareness in living systems, may help us find answers to one of the most profound mysteries in complex systems and science as a whole.

Syed Hussain Ather

Syed Hussain Ather(he/him) is a Ph.D. student in Medical Science at the University of Toronto where he performs research on the connectivity structures underlying schizophrenia. By applying dynamic causal modelling (DCM) to patient data, he searches for insights into spatiotemporal measures of connectivity across models of connectivity.His work, based on the free energy principle, has applications in neuroethics, epistemology, and AI. After completing his undergraduate in physics and philosophy at Indiana University-Bloomington, Hussain performed research in bioinformatics and computational neuroscience at the National Institutes of Health (NIH) before joining theKrembil Centre for Neuroinformatics(KCNI) in January, 2021. He also has interests in science communication and philosophy.