Good Science, Bad Philosophy: Predictive Processing as Reheated Kantianism

Some ongoing reflections on the turf war between enactivist cognitive science, predictive processing approaches, and Whiteheadian cosmology

Predictive Processing (PP) approaches (including Active Inference, the Free Energy Principle, etc.) are fantastic models that will surely continue to find important applications not only in cognitive neuroscience and developmental psychology, but in generative AI and robotics engineering. But when it comes to granting us insight into the ontology of mind and life, aside from making it clearer what living organization is not (a valuable service!), these approaches are in my humble opinion misguided and self-undermining. They reduce conscious organisms, including humans, to a predictive error minimization machines churning out Bayesian statistical calculations.

It may be that taking such a lens on the cognitive processes of living beings offers some degree of prediction and control over their behavior. It is already obvious that we can use this model to build impressively life- and mind-like machines. But when misplaced concreteness leads to the hypostatization of a partially descriptive model into a totalizing explanation of conscious agency, a helpful instrument becomes a metaphysical monster (no offense to monsters!). The attempt to over-extend PP approaches into ontology amounts to a kind of reheated Kantianism, often repackaged so as to be unrecognizable even to itself (hence overstepping its acknowledged epistemological value by unknowingly assuming an ontology). But it’s the same old phenomenal/noumenal divide under the Bayesian neurocomputationalist plastic wrap.

Here’s how The New Yorker described PP in their article on philosopher Andy Clark:

“Perception [is] not passive and objective but active and subjective. It [is], in a way, a brain-generated hallucination: one influenced by reality, but a hallucination nonetheless.

This top-down account of perception had, in fact, been around for more than two hundred years. Immanuel Kant suggested that the mind made sense of the complicated sensory world by means of innate mental concepts. And an account similar to predictive processing was proposed in the eighteen-sixties by the Prussian physicist Hermann von Helmholtz. When Helmholtz was a child, in Potsdam, he walked past a church and saw tiny figures standing in the the belfry; he thought they were dolls, and asked his mother to reach up and get them for him: he did not yet understand the the concept of distance, and how it made things look smaller. When he was older, his brain incorporated that knowledge into its unconscious understanding of the the world—into a set of expectations, or ‘priors,’ distilled from its experience—an understanding so basic that it became a lens through which he couldn’t help but see.”

Clark’s idea that perception is creative and cognition extended beyond the brain through the body and into the world is a redeeming interpretation of what PP entails. But in the end, he seems to revert to internalism by saying we only have access to our own neural states, while “the world itself [remains] off-limits” (“Whatever next? Predictive brains, situated agents, and the future of cognitive science” in Behavioral and Brain Sciences, 36(3) [2013], p. 183). Another PP theorist, Jakob Hohwy, is more consistent in denying that the mind is in any way in touch with the world. On his reading, “the mind [is] inferentially secluded from the world…more neurocentrically skull-bound than embodied or extended” (“The self-evidencing brain” in Noûs, 50(2) [2016], p. 259).

Karl Friston’s Free Energy Principle is a neat formal generalization of PP approaches so as to include all life and perhaps even all enduring physical entities in its scope. Tim Jackson and I have discussed the FEP frequently. Friston’s emphasis is definitely on refining the mathematical accuracy of the formal statistical model and finding useful applications for it, with the question of various philosophical interpretations ranging from internalist to enactivist put to the side.

According to FEP, biological systems resist environmental entropy by minimizing free energy. Evolution by natural selection has produced organisms whose perception-action loops are kept closely correlated with sensory inputs, with the goal of avoiding erroneous predictions. While the model is deployed in an instrumentalist way, already baked into the formal description is the bifurcation between internal and external states and the related distinction between valid and erroneous predictions. Just as with Kant’s account of a merely apparent world-for-me and a real but unreachable thing-in-itself, it is difficult to justify the distinction between valid and erroneous perceptions when all we have access to are sensations arising in our experience from unknown causes. In other words, upon what basis are we to distinguish between our perceptual guesses and our sensory impressions? There’s the assumption that sense data come from the outside world and represent external states. But isn’t our concept of externality just another confabulated error minimizing belief? The are similar problems with Locke’s division between primary and secondary characteristics, which Whitehead roundly rejects in The Concept of Nature (1920).

Despite the caution and instrumental attitude he professes in much of his published academic work, in popular interviews Friston often cannot help but wax philosophical on the implications of the FEP model. He refers to the brain as quite literally “a fantastic organ”: “our percepts are the results of fantasies and guesses and hypotheses that are either confirmed or disconfirmed” (YouTube interview 1). Friston even goes so far as to suggest the sort of ethical calculus that would spill out of FEP: “select those actions that minimize your uncertainty [and] provide the greatest information gain…[this] is often cast in terms of ‘optimal Bayesian design’” (YouTube interview 2). I grant this attempt at a mathematization of action is more useful than Kant’s categorical imperative. But both principles strike me has highly artificial renderings of how humans and all living beings actually realize relevance, make decisions, and take action. I’m reminded of Whitehead’s quip a century ago about how “Some people express themselves as though bodies, brains, and nerves were the only real things in an entirely imaginary world” (Science and the Modern World, p. 91).

…

In Book XI of The Prelude, as if speaking directly to Kant, William Wordsworth pays homage to the “animation and…deeper sway” evident in our participatory perceptions of the living world, while warning against the “narrow estimates of things” resulting from an exclusively critical or rational approach:

Suffice it here
To hint that danger cannot but attend
Upon a Function rather proud to be
The enemy of falsehood, than the friend
Of truth—to sit in judgment than to feel.

While for Kant, “the world emerges from the subject,” for Whitehead, “the subject emerges from the world” (Process and Reality, p. 172). Whitehead thus inverts the Kantian conception of subjectivity, such that the order and meaning of our experience is originally given to us by the order and meaning of the surrounding cosmos. “[The subject] is not productive of the ordered world, but derivative from it” (Process and Reality, p. 113).

PP models would apear to view the mind as fundamentally conservative and averse not only to direct contact with extra-neural causal flows but to the ingression of novelty. The model paints cognition as driven primarily by a fear of error rather than a love of truth (not to mention beauty and goodness). Cognition is depicted as a process of minimizing prediction errors to avoid surprise. As we’ve seen, this echoes the Kantian dichotomy where we never access the world directly but only experience our mind’s or brain’s internal models (self-models, world-models, etc.). It is simulacra all the way down—except, of course, for the scientific knowledge we have of the validity of this particular formal model.

Thus PP remains saddled with all the intractable problems intrinsic to Descartes’ good old fashioned representationalist theory of perception, leading straight to the epistemological skepticism that Kant and Hume were grappling with centuries ago. It’s the same old noumena dressed in computational drag. Many modern philosophers and contemporary PP-influenced cognitive neuroscientists alike “refuse to admit experience, naked and unashamed, devoid of their a priori fig leaf,” as Whitehead puts it (Process and Reality, p. 146). Whitehead develops his alternative theory of perception in an effort to reopen direct lines of communication between the reality of causal efficacy in nature and its various modes of appearance in our bodily life, sense perception, and intellectual feelings (also all part of nature). He also offers an explanation for the possibility of erroneous perceptions.

So, in summary, while PP has proven invaluable for the construction of powerful models, I find the ontological overreach of some of its proponents quite problematic. Needless to say, reducing conscious agency and cognition to a mechanistic process of error minimization flattens human experience. The brain is not a computer and does not locally produce consciousness. The FEP’s tensor equations are said to abstractly encode the relationships between the system’s internal states (its beliefs) and sensory inputs, mapping these through a network of inferences that allow the system to minimze the discrepancy between predicted and actual sensory data. But we should not make the mistake of imagining the brain is actually in the business of solving tensor equations and calculating Bayesian statistics (except, of course, when FEP theorists and physicists direct their imaginations toward the solution of said tensor equations)!

Whatever else cognition and consciousness may be, they depend not only on a functioning brain but on bodily life and its biosociotechnical couplings with its environment both synchronically and historically.

This may sound a lot like a defense of the enactivist approach. I have certainly been deeply influenced by that ecology of ideas, going back to my reading of Varela, Thompson, and Rosch’s The Embodied Mind: Cognitive Science and Human Experience (1991/2017) as an undergraduate. While most of its proponents remain phenomenologically oriented and tend to bracket explicitly metaphysical positions as unnecessarily speculative, I’ve long thought that the implicit ontology of enactivism must be a kind of process-relational panexperientialism. Evan Thompson has found himself arguing for some kind of post-physicalist neutral monist perspective, positing an underlying nondual something or other that is neither mind nor matter as traditionally conceived. But there’s a case for pushing beyond this to fully embrace panexperientialism. This shift moves us toward seeing consciousness not as the exclusive property of humans or animals but as a natural consequence of the way originally unconscious but already feeling-imbued energy flows tend to organize themselves.

…

So how do we avoid deluding ourselves with nice-sounding speculative ideas? It is important to remember that our ontological speculations arise from processes of abductive reasoning (in C. S. Peirce’s sense). Thus the role and method of speculative philosophy is not to launch into stating first principles and deducing consequences. Philosophy is the always tentative and revisable search for premises. FEP theorists are well aware that their model always assumes a reference frame. Ontological premises are what allow us to give sense to the concept of a reference frame in the first place. I don’t mean to suggest that scientific modelers must bow down to their philosophical masters. Philosophers have much to learn from the special sciences, and in some sense Whitehead’s cosmological scheme is just an attempt to generalize across disciplines ranging from electrodynamics, to physiology, to psychology, while also bringing scientific knowledge into congruence with our aesthetic and religious experiences.

My approach to speculative philosophy remains pluralistic in practice. I would never say, “This is The One True Onto-Logic to rule them all, I has it!” (Gollum voice). But in order to know what we mean when we compare frames and formalize tensor equations, we must already have at least a working definition of framing. Such a definition is not just another framing problem: it is our proximal zone of contact with The Frame, ie, with the process of realization itself. And of course, our references to That (or is it Thou?) must be made ironically, metaphorically, analogically, mythopoetically… But we reference It in all our formal models and research programs one way or another, whether we imply it or explicitly deny it (lest we forget that science assumes the intelligibility of nature). Here is how Whitehead frames the issue in his introductory chapter explaining his methodology in Process and Reality (p. 11-13):

“…every proposition refers to a universe exhibiting some general systematic metaphysical character. Apart from this background, the separate entities which go to form the proposition, and the proposition as a whole, are without determinate character. Nothing has been de­fined, because every definite entity requires a systematic universe to supply its requisite status. Thus every proposition proposing a fact must, in its complete analysis, propose the general character of the universe required for that fact. There are no self-sustained facts, floating in nonentity. … A proposition can embody partial truth because it only demands a certain type of systematic environment, which is presupposed in its meaning. It does not refer to the universe in all its detail.

One practical aim of metaphysics is the accurate analysis of propositions; not merely of metaphysical propositions, but of quite ordinary propositions such as ‘There is beef for dinner today,’ and ‘Socrates is mortal.’ The one genus of facts which constitutes the field of some special science requires some common metaphysical presupposition respecting the universe. It is merely credulous to accept verbal phrases as adequate statements of propositions. The distinction between verbal phrases and complete propo­sitions is one of the reasons why the logicians’ rigid alternative, ‘true or false,’ is so largely irrelevant for the pursuit of knowledge.

The excessive trust in linguistic phrases has been the well-known reason vitiating so much of the philosophy and physics among the Greeks and among the mediaeval thinkers who continued the Greek traditions. For example John Stuart Mill writes:

‘They [the Greeks] had great difficulty in distinguishing between things which their language confounded, or in putting mentally to­gether things which it distinguished and could hardly combine the objects in nature into any classes but those which were made for them by the popular phrases of their own country; or at least could not help fancying those classes to be natural, and all others arbitrary and artificial. Accordingly, scientific investigation among the Greek schools of speculation and their followers in the Middle Ages, was little more than a mere sifting and analysing of the notions at­tached to common language. They thought that by determining the meaning of words they could become acquainted with facts’ (Logic, Book V, Ch. III).

Mill then proceeds to quote from Whewell (History of the Inductive Sciences) a paragraph illustrating the same weakness of Greek thought.

But neither Mill, nor Whewell, tracks this difficulty about language down to its sources. They both presuppose that language does enunciate well-defined propositions. This is quite untrue. Language is thoroughly in­determinate, by reason of the fact that every occurrence presupposes some systematic type of environment.

For example, the word ‘Socrates,’ referring to the philosopher, in one sentence may stand for an entity presupposing a more closely defined back­ground than the word ‘Socrates,’ with the same reference, in another sen­tence. The word ‘mortal’ affords an analogous possibility. A precise lan­guage must await a completed metaphysical knowledge.

The technical language of philosophy represents attempts of various schools of thought to obtain explicit expression of general ideas pre­supposed by the facts of experience. It follows that any novelty in meta­physical doctrines exhibits some measure of disagreement with statements of the facts to be found in current philosophical literature. The extent of disagreement measures the extent of metaphysical divergence. It is, there­fore, no valid criticism on one metaphysical school to point out that its doctrines do not follow from the verbal expression of the facts accepted by another school. The whole contention is that the doctrines in question supply a closer approach to fully expressed propositions.

The truth itself is nothing else than how the composite natures of the organic actualities of the world obtain adequate representation in the divine nature. Such representations compose the ‘consequent nature’ of God, which evolves in its relationship to the evolving world without derogation to the eternal completion of its primordial conceptual nature. …[T]here can be no determinate truth, correlating impartially the partial experiences of many actual entities, apart from one actual entity to which it can be referred.”

…

Timothy Jackson

and I are meeting to discuss these issues in a few days, with a special focus on the examples of flow and improvisational performance. I am sure he will disabuse me of any exaggerations in the above critical treatment of PP approaches. I remain quite interested in the exciting applications of FEP and Active Inference. But if the literature is any indication, determining the compatibility of such models with enactivism and process-relational ontology remains very much a matter of debate.

Shaun Gallagher (with whom I took courses as an undergraduate) is among the critics of PP, arguing for its incompatibility with enactivism’s explanatory strategies. In his article, “Surprise! Why enactivism and predictive processing are parting ways” (in Possibility Studies & Society, Vol. 1(3) [2023]), he tries to lay out why improvisation proves difficult to explain in the PP framework, while such an account flows naturally from enactivism’s emphasis on brain-body-environment couplings. I know Tim has problems with both his portrayals of PP and improvisation, so we’ll spend some time digging in to where Gallagher may go wrong in what he denies (even if he is right in what he affirms).

Darius Parvizi-Wayne and colleagues argue in a recent article (“Forgetting ourselves in flow: an active inference account of flow states and how we experience ourselves within them” in Front. Psychol. 15:1354719 [2024]) that flow states, including improvisation, can in fact be addressed rather straightforwardly by PP models as (on my summary) the going offline of our self-model and the attunement of our implicit embodied understanding to environmental affordances. This idea extends to improvisational contexts, where performers do not rely on pre-set plans but instead adapt fluidly to the unfolding situation. Shutting off or dialing down our self-model allows for a kind of distributed attention—the improviser becomes attuned to the dynamic interplay between their body, the environment, and other performers (even if the body, the environment, and the other performers ultimately remain hallucinated inferences based on sense data we believe is likely to refer to to external states).

Brett Anderson, Mark Miller, and John Vervaeke jump into the fray with their paper “Predictive Processing and relevance realization: exploring convergent solutions to the frame problem” (in Phenomenology and the Cognitive Sciences [August 2022]). They argue that both frameworks, though distinct in their conceptual backgrounds, point toward a similar solution. PP leans on the idea of precision-weighting, assigning more weight to sensory inputs that prove more relevant (ie, more reliable relative to prior predictions). Such weighting shifts “the relative influence of bottom-up prediction errors and top-down predictions on our perceptions and actions,” as the authors put it. Relevance realization grows out of dynamical systems theory and focuses on how organisms balance competing goals (e.g., exploration vs. exploitation; specialization-generalization, focusing-diversifying, etc.) to zero in on relevant aspects of the world. The idea is that the trade-offs described by RR are also present in PP’s precision-weighting mechanisms. The authors hope that this convergence will allow for empirical demonstration of the RR framework by leaning on formal PP models to design a research program.

They also build on some of Anderson’s earlier work on the autism-schizotypy continuum and its relation to differences in the weighting of sense data. I found this the most interesting part of the paper, so it might be worth unpacking.

Individuals on the autism spectrum are often described in PP literature as deprioritizing higher-order predictions, meaning they rely more heavily on raw sense data. This could mean that they place less emphasis on contextual expectations and more on immediate, detailed information from the environment. As a result, they may have difficulty integrating broader contextual cues or making general inferences about unexpected social situations. On the opposite, schizotypy end of the continuum, individuals might exhibit over-weighting of high-level predictions, meaning that they rely too heavily on internal models or expectations, even when they don’t align with sense data. This could lead to hallucinations or delusions, where strong internal beliefs override incoming sensory information. The autism-schizotypy continuum thus reflects a variation in how different individuals weight top-down predictions versus bottom-up sensory data.

This example shows the usefulness of PP models for describing divergences from “normal” neural organization, but I still wonder if it gives us an adequate grasp of the conscious agency of human beings or living organisms more generally. To my mind, there’s nothing in such descriptions that is even suggestive of an explanation for consciousness, cognition, or organismic agency.

It is not uncommon for scientists to respond to philosophical gadflies like me by saying that all our ontological speculation is itself just more modeling. But I worry this attitude is precisely what leads to the implicit collapse of model into ontology, or science into philosophy (ie, as scientism). Ontology tells you what is available to be modeled. It’s the recipe and ingredients you need to bake various kinds of models. Models allow us to gain an instrumental grip on some deliberately isolated aspect of the world, and to discover invariances across various domains of application. Ontology concerns what we take ourselves to be modeling, to be perceiving and acting upon, as well as what we are as model-makers, as conscious agents. These are the two mysterious realities—the thinking I and the self-organizing cosmos (ie, the transcendental unity of apperception and of the system of nature)—that Kant punted into the Great Beyond.

…

In The Function of Reason (1929), Whitehead celebrates Darwin’s theory of natural selection as one of the great generalizations in the history of science. But he rejects the overextension of Darwin’s account of speciation into an explanation of life itself. Living beings are not simply survival mechanisms. If survival was the only aim, there is every reason for life to remain simple. Survival of the fittest (admittedly not Darwin’s phrase) gives us no explanation for the fact which stares us in the face: the evolution of more complex organisms comparatively deficient in survival power.

Whitehead’s book asks “What is the function of reason?” His answer is that reason is not unique to human heads but pervades the universe, even if its effects appear subtle and diffuse at inorganic scales. Reason is the art of life. And life can only be fully understood in terms of its threefold urge: to live (survival), to live better, and to live well.

Whitehead questions whether mechanistic science’s exclusive emphasis upon entropy does justice to the again self-evident counter-tendency of cosmogenesis. While today the stars and galaxies appear mostly to be wasting away, and while fossils of extinct species litter the earth, the upward trend of evolution has equally to be explained. Conscious agents devoted to scientific truth have emerged from the debris of dead stars. Aim is latent in the nature of things. Immanent teloi seed our sojourn into evermore intensely intimate forms of experiential togetherness.

Life cannot be reduced to the task of error minimization. Life is also the expression of Eros maximization.