Predictive Processing and The Problem of Demarcation

An objection often raised towards the predictive processing framework in theoretical neuroscience (henceforth PP) is that it is unfalsifiable (Hohwy, 2020, Colombo, Elkin & Hartmann, 2018). Insofar as predictive processing theories are taken to include or rely on the (unfalsifiable) Free Energy Principle, it appears they  would not meet Karl Popper’s (1969) criteria for demarcating what counts as science. In this essay I will introduce  the PP framework and the Free Energy Principle (Friston, 2010). I will then outline Popper’s (1969) theory of demarcation, and his notion of falsificationism. From there I will contrast PP with Darwin’s (1859) theory of natural selection to demonstrate that a framework need not be falsifiable to be considered scientific. I will conclude by arguing that the parallels between PP and evolutionary theory, the explanatory range of the PP research program, and its expanding capability to generate testable predictions results in it being, on balance, scientific in nature.

Predictive Processing and the Free Energy Principle 

I will start by outlining PP and the Free Energy Principle (henceforth FEP). PP is a leading research program in neuroscience and philosophy of mind. It can be considered a ‘toolbox’ of related theories (Hohwy, 2020). The general scheme is known as predictive processing (Clark, 2013; Clark, 2016), or prediction error minimisation (Hohwy, 2013). More specific, associated PP process theories are predictive coding (Kilner, Friston & Frith, 2007), and active inference (Friston, Mattout, & Kilner, 2011).These theories are sometimes, but not always, taken to be underpinned by the FEP, which is a mathematical, conceptual principle developed by theoretical neurobiologist Karl Friston (2010). 

The scope of this essay doesn’t allow for an expanded account of PP or its associated theories. The very basic idea is that perception and action are both active (not passive) processes; we constantly predict what we are going to experience next. We are driven by a goal of minimising prediction errors, which are generated when what we experience doesn’t match what we predicted, which then allows us to update our beliefs and generate better hypotheses about what will happen next. The aim is to have the most accurate picture of the world as possible. Our beliefs about the world have kept us alive so far, so it's better if we don’t find ourselves in a situation with sensory experiences that are surprising (like being exposed to extreme temperatures or being submerged in water for too long).That’s the extremely basic PP picture, and it is fundamental to all of the PP theories.

What is so interesting about PP is that, once fleshed out, this story about cognition promises to provide explanations about how an incredibly large and diverse range of cognitive phenomena are operationalised in the brain. These include consciousness, language, reading, emotion, dreaming, mental disorders, attention, thought, interoception, and our sense of agency (Hohwy, 2020).The field is still relatively new, but because PP has been able to demonstrate explanatory promise across such a broad range of areas, it has attracted a remarkable amount of attention, as scientists of various stripes, along with mathematicians and philosophers, explore its potential. 

The Different Kinds of Theories Under Predictive Processing 

What is of importance here, though, is to note the different kinds of theories that fall under the PP umbrella. The FEP is a principle, and it can’t be falsified. The principle “states that all systems that minimize their free energy resist a tendency to physical disintegration” (Colombo & Wright, 2018). In other words, all systems that do not disintegrate (like a drop of ink in a glass of water would, for example) abide by this principle to maintain their status as a self organising system. 

While the FEP doesn’t need to be invoked to work with PP and develop PP accounts, it provides crucial background information that can be used in predictive coding and active inference, both process theories that can generate predictions, which, in some cases, can be tested. Currently, precision-weighted predictive coding is leading the way in producing empirically verifiable theories (Hohwy, 2020). These hypotheses are typically based on minimal assumptions about certain cognitive processes that nonetheless have the potential to cast doubt on or falsify at least certain elements of the PP picture, or features of certain theories. This work includes (but is not limited to) research on cortical responses (Friston, 2018), ascending prediction error (den Ouden et al., 2009, Chennu et al., 2016), predictions modulating low level sensory activity (Kok et al., 2017), attention hiérarchies (Gordon et al., 2018), EEG N170 signals and prediction error (Johnston et al., 2017, Robinson et al., 2018) and sensory atténuation (Vasser et al., 2019). This is a tiny sample of the work that has been conducted on empirically verifiable PP theories, and the trend is expected to continue: “There is now a concerted effort in cognitive neuroscience to generate and test distinctive predictions of PP” (Hohwy, 2020, p.12)

Something that will become important in the following sections is that this research, as well as the predictive coding theory that underpins it, doesn’t necessarily need or depend upon the unfalsifiable FEP. But, the FEP provides a clear explanation of exactly why these processes unfold as they do, and, the mathematical formulations behind the FEP provide full accounts of how the mathematics behind predictive coding (and active inference) work (Friston, 2019). While the truth of the FEP is not our concern here, it is interesting (in the context of the scientific status of PP) that the success of predictive coding and active inference theories, in virtue of their interlocking mathematical and conceptual structures, provide support for the principle. 

Popper, Demarcation and Falsificationism

With an understanding of PP in hand, we can turn to an incredibly influential theory in the philosophy of science that would seem to entail that PP (insofar as it includes or relies on the unfalsifiable FEP), is unscientific -- Popper’s (1963) theory of demarcation. Of primary concern to Popper (1963) was working out how we could separate scientific from non scientific theories. This was an important undertaking, in the sense that there was a risk of theories being categorised as ‘science’ (such as Marx’s historical materialism) that, if publicly granted ‘scientific’ status, would take on a certain kind of authority that may not be warranted.(Godfrey-Smith, 2003). They would perhaps be taken as demonstrating ‘facts of the matter’, when there was good reason to be sceptical that they were dealing in facts. So, with the aim of demarcating science from not-science, or pseudoscience, Popper proposed an idea that was “simple, clear and striking” (Godfrey-Smith, 2003, p.57). What counts, he argued, is whether or not the theory is falsifiable. Stated formally, Popper argued that: “a hypothesis is scientific if and only if it has the potential to be refuted by some possible observation” (Godfrey-Smith, 2003, p.58) In other words, it must be the case that the theory generates hypotheses that can be tested such that they could demonstrate the theory to be false. We can make claims, he thought, and we can gather evidence, but we cannot ever be properly sure of the status of a theory until we have falsified it. And then, all we can be sure of is that it is not true and we need a better theory.(Godfrey-Smith, 2003). The accompanying claim (that we can never confirm a theory, only disconfirm one), is a controversial matter that will not be addressed here. What is of importance is the declaration that if something cannot be tested and falsified, it should not be considered scientific. 

The benefits of such a theory are obvious.It allowed for a clear separation between theories from established sciences such as physics and biology, and somewhat more (scientifically) questionable programs (Popper was concerned with psychoanalysis, for instance). Popper’s falsificationism was an incredibly influential idea that still carries a lot of weight today (Godfrey-Smith (2003). Of course, if we want to grant that Popper is right, then this entails that the FEP, and any associated PP theory that relies on it, is unscientific. This is a controversial statement, though. PP is a program that sits, roughly, in the field of theoretical neuroscience (although it spans quite a diverse range of fields including biology and machine learning). People who conduct research under PP include cognitive scientists, physicists, biologists and computational neuroscientists. It is employed in explicitly scientific fields like neurophysiology. We can see, now, that PP is associated with an unfalsifiable, tautological principle, has strong explanatory power, has sub theories that can provide testable hypotheses, and is employed by scientists. Each one of these traits accompanied another theory that didn’t fit nicely with Popper’s account -- Darwin’s (1859) theory of natural selection. 

Popper and the Theory of Natural Selection

Popper is well known for claiming that the theory of natural selection (henceforth NS) is a tautology and unfalsifiable and thus not scientific,and then apparently changing his mind (Sober & Elgin, 2017). His theory of demarcation had trouble dealing with Darwin’s theory of evolution, in that it was unfalsifiable, but to many people (including the scientists in the field) it was perfectly obvious that it was a scientific program. Popper’s conclusion notably conflicted with the claims of “distinguished evolutionary biologists C. H. Waddington, G. G. Simpson, J. B. S. Haldane, and R. A. Fisher” (Sober & Elgin, 2017, p.34). The following set of quotes from Popper follow his journey with his opinion of whether or not NS is a science. It is relevant to us because of the similarities between NS and the FEP. 

His view began with “I see in modern Darwinism the most successful explanation of the relevant facts” (Popper, 1957, p. 106) This is relevant because PP is often considered to provide the best explanation for a wide range of facts about cognition. (Hohwy, 2020, Friston, 2019), and thus roughly fits this broad claim.

This was followed the next year by the statement that “I have come to the conclusion that Darwinism is not a testable scientific theory,but a metaphysical research program -- a possible framework for testable scientific theories.” (Popper, 1976, p.168). This is relevant because PP could definitely be considered a research program, and the claim that the FEP -- a mathematical, theoretical construct really is what drives physical processes could be considered a metaphysical claim. We’re also getting closer now, in that, as Popper is allowing for NS, PP is a framework for testable scientific theories (such as the aforementioned work in predictive coding).

Five years later, Popper claimed that “It does appear that some people think that I denied scientific character to the historical sciences, such as palaeontology, or the history of the evolution of life on Earth. This is a mistake, and I here wish to affirm that these and other historical sciences have in my opinion scientific character; their hypotheses can in many cases be tested..” (Popper, 1981, p. 611) The question here is would Popper consider PP to have ‘scientific character’? What seems of importance to him in this case is that hypotheses generated by a theory can in many cases be tested. I don’t think PP meets this criteria. Despite the remarkably wide engagement with PP, there is actually only a relatively small body of empirical evidence and testable theories. The FEP, while not necessarily required for a PP account of ‘x’, is very much wrapped up in the larger story. And there is charge of being tautological that can levelled at much of PP: because what produces an action according to PP is a prior belief combined with new evidence (or likelihood), that means that you can say that any action conforms to PP, given the right prior and likelihood (Hohwy, 2020). PP as a field is quite young, though. While still controversial, it’s explanatory potential is so great scientists all over the world, in different fields, working on diverse research programs are trying to generate testable predictions to further legitimise it. It appears at this point, though, that it’s not close enough to NS to count as having scientific character, on Popper’s account. 

There is still a big win here though, for PP, in that it turned out that Darwin’s theory of evolution being falsifiable, in the end, didn’t mean (apparently, for Popper) it didn’t have scientific character. Another common objection to PP, and particularly the FEP, is that its proposed scope (it is often considered a unified theory of the mind, running all cognition, and the FEP has been claimed by some to underpin all living systems), means that it is a preposterous idea (Colombo & Wright, 2017; Colombo, Elkin, & Hartmann, 2018). But if we refer back to NS, we can see that “the issue whether the free energy principle is preposterous cannot be decided just by pointing to its explanatory ambition, since this would also invalidate the theory of evolution” (Hohwy, 2015, p. 10)

Concluding Remarks 

So far, I have explained PP and the FEP, and demonstrated that under the PP ‘umbrella’, or in the ‘toolbox’, are different kinds of theories. The FEP, as a principle, is not falsifiable. It would therefore not meet Popper’s criteria for being scientific. And it is true that while PP researchers don’t need to subscribe to the FEP, many do, and believe that it is perfectly obvious that the explanatory range of the FEP, and its mathematical fit with empirical PP process theories, point to its truth. I looked at parallels between PP and the theory of natural selection, noting similarities in the context of what counts as science. While it seems PP still wouldn’t meet Popper’s criteria, NS gives us two strong responses to the main objections levelled at PP: that is unfalsifiable and has preposterously ambitious scope. Neither can decide PPs fate, though, without invalidating the theory of evolution. 

The PP framework is generating testable hypotheses, as was demonstrated with predictive coding research. It seems that we have a framework that cannot be denied scientific character without denying the same to NS. We can still argue that PP doesn’t have very much empirical evidence, and it hasn’t yet generated very many testable predictions. But it also seems clear that we have a new kind of framework on our hands. PP is like nothing that has come before it. It can be analysed at levels including philosophical, biological, mathematical, and programmable. It has ‘take it or leave it’ features (like the FEP, which is controversial, but is a powerful and fascinating concept if you do give the PP picture weight) I don’t see good enough reason to deny that PP is a science. Indeed, “PP is demonstrating significant potential for advancing several important debates in cognitive science and on to philosophy of biology” (Hohwy, 2020, p.10). If PP doesn’t fit neatly into certain boundaries, perhaps we may wish to redraw them. PP isn’t going away, and I suspect it will keep philosophers of science busy for a while. If we want to categorise it one way or another now, then is it, on balance, scientific. 

Bibliography 

Allen, M. (2018). The foundation: Mechanism, prediction, and falsification in Bayesian enactivism. Comment on" Answering Schrödinger's question: A free-energy formulation" by Maxwell James Désormeau Ramstead et al. PhLRv, 24, 17-20.

Colombo, M., Irvine, E., & Stapleton, M. (Eds.). (2019). Andy Clark and his critics. Oxford University Press, USA.

Colombo, M., Elkin, L., & Hartmann, S. (2018). Being realist about Bayes, and the predictive processing theory of mind. The British Journal for the Philosophy of Science

Colombo, M., & Wright, C. (2018). First principles in the life sciences: the free-energy principle, organicism, and mechanism. Synthese, 1-26.

Colombo, M., & Wright, C. (2017). Explanatory pluralism: An unrewarding prediction error for free energy theorists. Brain and Cognition, 112, 3-12.

Darwin, C. (1964). On the origin of species: A facsimile of the first edition. Harvard University Press.

Godfrey-Smith (2003) Theory and Reality: An Introduction to the Philosophy of Science, University of Chicago Press

Hohwy, J. (2020). New directions in predictive processing. Mind & Language, 35(2), 209-223.

Hohwy, J. (2014). The neural organ explains the mind. Open MIND. Frankfurt am Main: MIND Group.

Friston, K. (2010). The free-energy principle: a unified brain theory?. Nature reviews neuroscience, 11(2), 127-138.

Friston, K., Mattout, J., & Kilner, J. (2011). Action understanding and active inference. Biological cybernetics, 104(1-2), 137-160.

Friston, K. (2018). Does predictive coding have a future?. Nature neuroscience, 21(8), 1019-1021.

Kilner, J. M., Friston, K. J., & Frith, C. D. (2007). Predictive coding: an account of the mirror neuron system. Cognitive processing, 8(3), 159-166.

Popper, K. (1959) The Logic of Scientific Discovery.

Popper, K. (1963) Conjectures and Refutations, Routledge

Sober, E., & Elgin, M. (2017). Popper’s Shifting Appraisal of Evolutionary Theory.

Wiese, W., & Metzinger, T. (2017). Vanilla PP for philosophers: A primer on predictive processing.