Minds as Predictive Engines: a Matter of Scope

Introduction

Predictive Processing (PP) is a framework within cognitive neuroscience that provides a computational account of the mind. It’s basic promise is a unifying explanation of how cognition, perception and action are related and driven. Roughly, the central idea is that brains are multi-layered prediction engines that measure expected sensory stimuli against incoming sensory stimuli, and then act to minimise any discrepancy between the two. 

In this paper, I will provide a thorough outline of PP. Then, I will present the key arguments for this account of the mind. From there, I will explain how the explanatory scope of PP can be understood in different formulations, ranging from broad and uncontroversial to strong. This will lead us to the central arguments against PP. Finally, I will conclude that while stronger accounts of PP are vulnerable to considerable objections, the basic framework of PP has sound conceptual promise and exceptional heuristic value, and therefore warrants further empirical investigation. 

The Process 

I shall begin by expanding on the above definition of PP. On this account, brains generate internal models of reality, which they then compare with incoming data. These models are based on prior beliefs and experiences. Brains generate multiple hypotheses about the sources of incoming sensory data, and the most likely hypothesis becomes a prediction. To illustrate, imagine seeing what looks like a human being with wings walking down the street. On the PP account, your brain will use prior beliefs and experiences to generate the most likely hypothesis about what you are seeing. This could be that the person is wearing a costume. The brain then compares the generated prediction with incoming sensory stimuli, identifies any errors between the two and updates its internal models, so that it can predict and therefore perceive more accurately moving forward. In reality, this could look like you stopping and intentionally directing your attention towards the person to perceive if it is, indeed a costume they are wearing, so that you can confirm or update your internal models of how reality actually is. Errors are assigned precision weightings (how significant they are, given the context). This process is called prediction error minimisation (PEM). 

The Hierarchy

According to PP, the above computational process is ongoing and takes places in a multilayered brain. The idea is that predictions about sensory stimuli are generated and move in a top-down fashion, while actual sensory data comes in at the bottom and is analysed against the predictions. If they do not match, then a prediction error signal is sent both laterally and upwards through the layers, and the parameters at higher levels are updated in accordance with Bayes Theorem: a mathematical formula that updates the probability of a hypothesis being correct in light of new evidence (Metzinger & Wiese 2017, Sims 2017). These layers are typically understood to represent different temporal and spatial scales (Metzinger & Wiese 2017).

Action

The above explains how we perceive, but it’s also possible to explain how an organism acts by appealing to the same computational framework. According to Clarke (2016), PEM can also make our bodies move, pushing us to generate sensations so that our predictions can be confirmed or rejected, optimising the accuracy of our internal model. The point here is that the process of perception is not passive. It is constructed by action-oriented predictions determined by the kind of organism we are.

The Purpose

Now I’ve outlined how the process works, a natural question to address is why our brains would constantly seek to minimise prediction error. On the PP account, to determine what incoming signals are important amongst a field of constant sensory stimuli, our brains make well-informed guesses as to the causes of the incoming data (Clarke, 2016). Imagine being in a tight crowd of people at a demonstration that could turn violent. Your body is constantly receiving sensory input, in the form of competing voices and sounds, the people and objects you can see around you, and the physical touch of those near you. Your brain would have to make sense of that chaotic, sometimes uncertain sensory input to increase your chance of getting it right, and minimise risk to yourself. If PP is on the right track, our brains do this all the time, by making educated guesses (predictions) about what is causing our sensory input.

Arguments for PP 

I shall now turn to some of the best arguments for PP - considered one of the most fast-moving and cutting-edge areas of research in contemporary philosophy and the cognitive sciences (Kirchhoff 2018). It is gaining in influence, is increasingly utilised in theoretical and experimental studies and may plausibly dominate the field of neuroscience in the near future (Howhy 2013, 2016). The arguments for PP can be roughly split into three groups: conceptual potential, a general fit with structures of the brain and other mental phenomena, and experimental/empirical. I shall outline each in turn.

Conceptual 

If PP theorists are right, the account could: construct strong conceptual bridges between theoretical and empirical research on cognition; reveal unifying relationships between apparently unconnected mental processes, and provide a significantly unified picture of cognition, perception and action (Wiese & Metzinger 2017, Howhy 2013, Clarke 2016).

A General Fit

PP also fits well with general physiological and anatomical features, the plasticity, and the “functional segregation and connectivity” of the brain (Howhy 2013, p.8). Mental phenomena such as consciousness are also newly illuminated by the promise of a fit with PP. Hohwy (2013), makes the case that our conscious awareness is actually constituted by a combination of the most accurate predictions at each level of the cortical hierarchy outlined above. 

Experimental 

Clarke (2016), presents a number of suggestive experiments and illusions that appear to show the PP mechanism at work. One example taken from Remez et al (1981) and Remez & Rubin (1984), involves sine-wave speech – a degraded, incomprehensible version of a speech recording. Clarke encourages readers to try the experiment out. The idea is that you listen to the degraded version and likely find it unintelligible. Then, you listen to the original, clear recording. The next time you listen to the degraded recording, you should notice an improved ability to make out the words. In fact, when you hear the degraded version again, it’s almost impossible not to understand it - as in, you can’t hear it the way you originally did. (Clarke 2016, p. 55). The claim is that this is evidence of the brain generating more accurate predictions about incoming sensory input by measuring against its newfound knowledge. 

Empirical

Clarke also provides empirical evidence that seems to back up the PP framework, including “an emerging body of supportive fMRI and EEG work...that also reveals just the kind of relationships posited by the predictive processing story” (Clarke 2016, p. 44). In addition, there is “ample evidence” that top-down influences control how we process incoming sensory data by shaping thalamocortical networks to consistently generate predictions about future sensory stimuli (Engel et al. 2001, from Wiese & Metzinger 2017, p. 2). 

A Matter of Scope 

Before examining the arguments against the PP account of the mind, I shall provide context via the notion that one can endorse different strengths of PP. Sims (2017) suggests that there are different formulations of PP, ranging from broad to strong and controversial. He divides them into “four very general positions” (Sims 2017, p. 4), as below:

Sims (2017, p. 18)

Clarke, however, considers the Free Energy Principle (to be explained shortly), representative of the “maximal version” of PP (Clarke 2013, p. 187). For the purposes of this paper, I shall use Clarke’s definition; when I refer to maximal PP, I am referring to the FEP. 

FEP

The FEP is a principle that explains how biological systems maintain order by restricting themselves to a limited number of states. Friston describes it as: “...essentially a mathematical formulation of how adaptive systems (that is, biological agents, like animals or brains), resist a natural tendency to disorder.” (Friston 2010, p.) Friston ultimately claims that PP is underpinned by FEP in such a way that the underlying principle behind all biological processes is the drive to maintain an orderly state by minimising surprise, via PEM. In this way, Friston implies that in its strongest formulation, PP could provide a Grand Unified Theory of the mind - a principle that unifies all the sciences of the mind (Sims 2017). While ambitious, there is a clear link between PP and FEP, in that conceptually, the computational framework of PP and the statistical physics behind FEP seem to share analogous central mechanisms of minimising error, also known as surprise (Friston, 2010, Kirchhoff 2018). 

Arguments against PP 

Now FEP has been outlined, I shall examine two of the strongest objections to the PP account of the mind. The first is a conceptual concern known roughly as the dark room problem (Friston 2012, Clarke 2016, Klein 2018). The central idea is that if the underlying principle of the mind is to minimise prediction error (or surprise), then organisms should seek out environments where surprise is kept to an absolute minimum, such as a quiet, dark room. Clearly, this is not what organisms actually do, therefore, FEP seems misguided. Sims (2017), proposes that this objection can be overcome by accepting his above mixed account (the position that PP drives some, but not all cognitive processes), as this account doesn’t require PEM to be the actual underlying principle of all biological processes in the way that FEP does.

Another key objection is raised by Klein (2018) and Sims, who calls it “The Triviality Problem” (Sims 2017,p. 7). Both claim that PP in its strongest, FEP formulation is problematic, because, at this wide scope, it becomes a tautology that lacks sufficient explanatory power. Sims refers to what Friston himself wrote: “The tautology here is deliberate [...] Like adaptive fitness, the free-energy formulation is not a mechanism or magic recipe for life; it is just a characterization of biological systems that exist.” (Friston et al. 2012b, p.2 taken from Sims 2017, p.8). Klein (2018), argues that the tautological nature of this argument renders it explanatorily weak, in that it explains itself by stating a definition. He also makes the point that natural selection requires differing levels of fitness between organisms - some need to be performing better than others within a given niche (Klein 2018, p. 2553). He infers that if this kind of variation is necessary, then some organisms simply aren’t optimal, which implies PEM directed towards an optimal state is not the underlying principle of all biological systems. 

Conclusion

In this paper, I firstly outlined the PP account of the mind. From there, I examined the central arguments for this account, and presented Sims’ (2017) general formulations of PP, which turn on the account’s explanatory scope and culminate in the FEP. I then examined key arguments against PP.

At this point, the following can be logically inferred: Firstly, if Sims’ (2017) version of mixed PP is on the right track, it provides access to developing accounts of mental phenomena that incorporate the basic framework of PP without relying on FEP. Secondly, although FEP appears alluringly linked to PP, unless FEP theorists can successfully overcome both the dark room and the triviality objection, this position is problematic. Thirdly, the field has grown richer with conceptual arguments and impressive experiments; what is most required now is further empirical evidence for the neural correlates of the PP computational framework. And lastly, in the face of objections, PP still holds enormous promise. It has extraordinary potential to unify phenomena within organisms and theories across disciplines. Therefore, the field of predictive processing warrants significant further empirical research, particularly in the domain of developing mixed PP accounts that aim to map PP’s computational framework at the neuronal level, while avoiding the objections FEP is vulnerable to. 

Bibliography

Clark, A. (2013) Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences 36, 181-253.

Clark, A. (2016) Surfing Uncertainty: Prediction, Action, and the Embodied Mind. New York: Oxford University Press. 

Friston, K. (2010) The free-energy principle: a unified brain theory? Nature Reviews Neuroscience volume 11, p. 127–138

Hohwy, J. (2013) The Predictive Brain. Oxford: Oxford University Press.

Kirchhoff, M. (2018) Predictive brains and embodied, enactive cognition: an introduction to the special issue. In Synthese p. 2355-2366 https://link-springer-com.simsrad.net.ocs.mq.edu.au/journal/11229/195/6/page/1 Accessed 12 June 2018

Klein, C. (2018) What do predictive coders want? In Synthese June 2018, Volume 195, Issue 6, pp 2541–2557 https://link-springer-com.simsrad.net.ocs.mq.edu.au/journal/11229/195/6/page/1 Accessed 12 June 2018

Sims, A. (2017) The Problems with Prediction. [1, 1-18] In T. Metzinger & W. Wiese (eds.) Philosophy and Predictive Processing. Frankfurt am Main: MIND Group.

Wiese & Metzinger (2017) Vanilla PP for Philosophers: A Primer on Predictive Processing. [1, 1-18] In T. Metzinger & W. Wiese (eds.) Philosophy and Predictive Processing. Frankfurt am Main: MIND Group.