The Contents of Bayesian Perception

By: Alexander Miller Tate, MSc Mind Language and Embodied Cognition 2015, University of Edinburgh


This paper presents a novel defence of the Perceptibility Thesis (PT). PT is the highly controversial claim that the properties of which we can conceive coincide with those that we can perceive. In particular, PT suggests that we can perceive abstract properties such as vegetarianism and scariness which lack “…characteristic looks” (Prinz, 2006: 434), just as well as we can perceive concrete entities like dogs and tables.

After reviewing Prinz’s strategy for arguing for PT, this paper will adopt a different approach. A Hierarchical Bayesian Predictive Processing (HBPP) account of the mind, I argue, implies that there is no a priori limit on the ‘abstractness’ of the content that can be “perceptually embedded…part of the content of the phenomenology of perception.” (Levy, 2014: 344). This is not quite because perception is mediated by beliefs, but because belief and perception according to HBPP, differ only in relation to “…the time scale of the represented processes and their degree of invariance or perspective independence” (Hohwy & Rajan, 2012: 10). As a result, any content that can be represented in belief can also be similarly represented in perception, as they are realisations of the same processes of prediction-error minimisation. How this might occur in practice is expanded upon briefly, before the essay concludes.

Prinz on abstract percepts

I shall quickly define some of the terms crucial for understanding Prinz’s argument. By perceptual representation, or percept, he means a kind of mental representation that is formed by suitably transforming a signal from a sensory transducer. These in turn, very roughly, transform “raw” perceptual data from the environment, such as reflected light waves or vibrations in the air, into a corresponding brain signal. A perceptual representation thus understood can be thought of, loosely, as having the function of representing such input (Prinz, 2006: 436).

Perceptual pattern-matching is the key process by which we come to actually perceive anything on Prinz’s account. It is the process of ‘matching’ a perceptual representation (P) to a further stored representation (Q). This ‘matching’ need not involve any explicit process of comparing P and Q. Crucial is the idea that “…perceiving involves recognition.” (Prinz, 2006: 436), whether this is achieved by P being compared with Q, or P ‘triggering’ Q in some other sense. I perceive X on Prinz’s account only if X impinges on my senses, the input from corresponding transducers is transformed into a perceptual representation, P, and P is matched against a stored representation Q, which represents X.

With these concepts in place, Prinz argues for PT in the following way.

P1) Abstract concepts are “…associated with perceptual representations…in order to secure reference to properties in the world.” (2006: 450).

P2) When any such representation is ‘applied’ in perceptual pattern-matching, it confers the content of its associated concept/s.

  1. C) So, when a stored perceptual representation associated with an abstract concept, such as ‘uncle’, is matched with a perceptual representation constructed from perceived objects in the environment, this matched representation has the abstract content conferred to it.

Prinz argues for P1 by considering that abstract properties must somehow get to refer to the concrete particulars that we attribute them to. For instance, if I see a rabid dog in my visual field and consciously appraise it as ‘scary’, then it seems that my concept of ‘scariness’ gets to refer to an instance of a genuine property. It does this successfully, according to Prinz, “…because it is causally responsive to situations of this kind.” (2006: 448) There are things in the world which are scary, which lack surface uniformity. We may nevertheless gradually develop a concept of scariness by applying it to those things liable to cause an experience of fear in us. Our concept of ‘scariness’ is set up so as to be tokened in situations where something scares us. The concept represents what it does in virtue of being a reliable detector of tokens of a corresponding property (Prinz, 2006: 447). Since perceptual representations (via their connection with sensory input) are, on Prinz’s story, the only avenue by which things in the external world can cause anything to happen in the mind, the concept of scariness can only detect scary things in virtue of being caused by perceptual representations, themselves caused by scary things impinging upon our senses. So all of our concepts are associated with a collection of perceptual representations, in order to secure their reference to objects in the world.

Some cases may involve only passing associations (Prinz, 2006: 446). If I ask every vegetarian in the room to raise their hand, I only temporarily associate ‘raised hand’ percepts with the property of being a vegetarian. Nevertheless, such percepts temporarily become ‘reliable detectors’ of vegetarianism, securing that concept’s reference to the abstract property in the world.

This broadly Dretskean (Dretske, 1981) approach to how psychological states refer to things in the external world (their intentionality) is heavily implicated in Prinz’s argument for P2 (the claim that perceptual representations transfer the content of their associated concepts in perceptual pattern-matching) as well. Since representation “…is determined by detection and function…” (2006: 450), and percepts associated with concepts function to enable the organism to detect those properties in the world, those percepts, so the thought goes, count as representing the properties in question. So when I associate a stored ‘raised hand’ percept with the concept of vegetarianism and subsequently use those stored percepts to match a percept derived from objects impinging on my retinae, I literally perceive the property of vegetarianism. This occurs because the stored percepts are temporarily associated with the concept of vegetarianism, and thus function to detect the property in the world (i.e, represent it). In a case of less transient association, I understandably associate my concept of scariness with stored percepts of (among other things) rabid dogs. So when a perceptual representation caused by the impinging of a rabid dog upon my senses is matched against this percept, I literally perceive the associated concept of scariness. Again, this occurs because the stored percept serves to detect the property of scariness (and rabidity, and perhaps to a lesser extent dogs in general) in the world.

It is not enormously important in what follows whether Prinz’s argument is sound. But there are two features of it that we should note. Firstly, Prinz’s story is wedded closely to a very controversial account of how mental representations come to refer to things in the world (cf. Adams & Aizawa, 2010: §3&4). While some kind of ‘causal’ theory, as utilised above by Prinz in his arguments for P1 and P2, is a popular answer to this notoriously tricky question (cf. Jacob, 2014), this controversy should give us pause to wonder just how secure Prinz’s conclusion is. Relatedly, one might suggest that there is insufficient independent evidence for this story of how psychological states refer to the world for Prinz’s argument to count as anything other than a reductio of it. After all, Prinz goes to great lengths to acknowledge the counter-intuitive nature of PT (2006: 440). If this controversial account of intentionality strongly supports PT, then perhaps that is so much the worse for it. If we are already suspicious of PT, a relatively poorly independently supported theory of psychological reference is unlikely to convince us of its truth, or even plausibility.

Ideally, a defender of PT will want to argue from a theoretical perspective for which there exists much independent supporting evidence, both empirical and theoretical. This would diminish the threat posed by any reductio-style response to PT as, dialectically, it would appear much more ad hoc and unsubstantiated. I believe that HBPP can provide such a framework, though even if it turns out to be as controversial as the above, independent support for PT still seems worth seeking out.

HBPP and the continuity of belief and perception

HBPP, which I will explain in the next paragraph, is a movement in Cognitive Science that aims to provide a wide-ranging, high-level explanation of human brain activity, as well as cognitive performance more generally. There is mounting evidence in its favour across disciplines, it shows signs of being able to revolutionise cognitive science within the next few years (Hohwy, 2014; Clark, 2013), and it is already being deployed fruitfully in explanations of diverse and complex psychological and psychiatric phenomena, including but not limited to delusion and psychosis (Hohwy & Rajan, 2012; Corlett et al, 2007; Frith, 2005), depression (Gradin et al, 2011), the functioning of mirror neurons (Kilner, Friston & Frith, 2007) and binocular rivalry (Hohwy, Roepstorff & Friston, 2008). So it seems safe to say that if HBPP were found to support PT, then the latter thesis would inherit greater plausibility than that suggested by our pre-theoretic intuitions. In the outline of HBPP that follows, I draw heavily from Clark (2013).

One key insight brought by HBPP is that perception is the result of the brain’s attempts to predict its own sensory input. ‘Bottom-up’ information from the sensory transducers is compared with a ‘Top-down’ generative model of expected sensory input. Sensory information is carried into higher-processing as only when it conflicts with this generative model; if it what was expected, then it is discarded early in processing. In this context, the sensory information that proceeds further in the processing hierarchy is referred to as prediction error, as it conflicts with the brain’s prediction. Moreover, the brain is simultaneously predicting the expected precision of the information coming from the sense organs. This corresponds intuitively to the idea that we treat information from a given sensory modality as less reliable when we expect it to be of poor quality (for example, vision on a very foggy day). Using the information received in this way, the generative model is updated appropriately – to include the previously unpredicted signal if it is considered to be of sufficiently high precision, or not if it is judged more likely to be noise. The generative model predicts the input that is most probable, given a prior distribution and the weighted prediction error. The update is taken to be Bayes-optimal. That is, the model adopts the most probable ‘hypotheses’ (expected sensory input), given the ‘evidence’ (prediction-error) and its weighted precision. The overall ‘goal’ of the system is to minimise future prediction-error. What we perceive then is, in a sense, our brain’s ‘best guess’ about the state of those aspects of the world impinging upon our senses.

Importantly, the above is also taken to be true for belief formation and revision. Beliefs are treated as the brain’s best predictions about the expected state of the world, and are revised only in the presence of significant quantities of precise prediction-error.

In arguing for PT, this equivocation will prove crucial. Since according to HBPP both belief update and perception are realised by predictions, prediction-error and generative model update, the two processes are of the same basic kind. As Hohwy & Rajan put it,

“The difference between belief and perception lies in the time scale of the represented processes and their degree of invariance or perspective independence. There is no further special difference between them…” (2012: 10)

Two things should be noted about this result. Firstly, it is distinct from the claim that perception is ‘belief-mediated’, or that our beliefs can in some way influence what we perceive. This still treats percepts and beliefs as fundamentally different mental phenomena. HBPP goes one step further. It says that our beliefs are the same kind of thing as our percepts; the product of prediction-error minimisation. They are more long-lasting than percepts, and have a degree of independence from perspective, but exist along the same continuum.

Secondly, it has a consequence that is of significant interest to defenders of PT. Since both beliefs and percepts are realised by a generative model, we should have confidence that each will be able to represent the same content; that of the aforementioned model. So if we can believe things that utilise concepts referring to abstract properties, then we can in principle perceive those same properties. If we predict a certain state of the world, such as ‘the people with their hands raised are vegetarians’ or ‘a scary dog is in front of me’, and therein successfully minimise prediction-error, then that is precisely what we will perceive, abstract content and all.

This point is raised most clearly by Levy. As he puts it for the case of patients suffering from delusions of alien control,

“As a consequence [of its ability to account for the large quantities of falsely generated prediction-error in the motor system], the content of the supernatural agency models comes to be embedded in the perceptual experience. The delusional patient experiences her movement as controlled by outside forces; she does not interpret her movement as so controlled.” (Levy, 2014: 343, emphasis in original)

I am now in a position to explicitly render my argument for PT.

P1) According to HBPP, beliefs and percepts are realised by the same kind of brain process – attempts to minimise sensory prediction-error.

P2) As a consequence, content normally thought of as belonging to a belief can become embedded in an agent’s perceptual experience.

P3) This content can, in principle, contain abstract reference.

  1. C) Content referring to abstract properties can become embedded in perceptual experience, and thus literally perceived.

Consider again the case of me perceiving a rabid dog. My brain must adapt its generative model in light of the flurry of prediction-error generated when it appears in my visual field. Fortunately I have a prediction that makes quick work of the new, highly specific, evidence – ‘a scary dog is in front of me’. Since the precise content of that prediction is used in minimising perceptual prediction error, I literally perceive not only the shape, size and colour properties of the dog, but also the property of ‘scariness’ that it instantiates.

Relatedly, researchers have hypothesised that affective predictions are involved in the generation of visual experience in all (or nearly all) cases. On this view,

“…responses signalling an object’s salience, relevance or value do not occur as a separate step after the object is identified…affective responses support vision from the very moment that visual stimulation begins.” (Barrett & Bar, 2009: 1325)

If this view is correct, then it would appear that abstract content is involved in our perceptual predictions more often than not. We do not generally perceive objects simpliciter, but objects as instantiating certain affectively salient (oftentimes abstract) properties, such as scariness, or edibility.

Further, in the vegetarian example, what hypothesis best accounts for a number of raised hands all of a sudden impinging upon my visual field? Plausibly it is something like ‘the vegetarians in the room are responding to my request’. Since the property of vegetarianism is explicitly encoded in the prediction of my generative model, that content is embedded in the percept thus realised.


I have given an argument for PT from a very popular and fruitful working assumption in contemporary cognitive science; that the brain is a prediction-error minimiser involved in predicting its own sensory inputs. I have argued that this case has certain dialectical advantages over the argument provided by Prinz. Specifically, though HBPP is surely committed to the idea that mental goings-on represent the world in some sense, it is not committed to any account of how that comes about.

Since the generative model realises both beliefs and percepts, the content of one can become embedded in the other. Wherever predictive content invokes abstract concepts to make sense of some incoming sensory prediction error, those very abstracta can come to be embedded in the agent’s perceptual experience. Pace Prinz, if we accept HBPP, then we have good reason to think that perception is not, in principle, significantly outstripped by conception.




Adams, F. & Aizawa, K. (2010), “Causal theories of mental content”, in Zalta, E.N. (ed.) The Stanford Encyclopedia of Philosophy, available at, retrieved 31/03/2015

Allen, C. (2009), “Teleological notions in Biology”, in Zalta, E.N. (ed.) The Stanford Encyclopedia of Philosophy, available at, retrieved 31/03/2015

Barrett, L.F. & Bar, M. (2009), “See it with feeling: affective predictions during object perception”, Philosophical Transactions of the Royal Society: Biological Sciences (364), pp.1325-1334

Clark, A. (2013), “Whatever next? Predictive brains, situated agents, and the future of cognitive science”, Behavioral and Brain Sciences (36:3), pp.1-73

Corlett, P.R., Murray, G.K., Honey, G.D., Aitken, M.R.F., Shanks, D.R., Robbins, T.W., Bullmore, E.T., Dickinson, A. & Fletcher, P.C. (2007), “Disrupted prediction-error signal in psychosis: evidence for an associative account of delusions”, Brain, 130, pp.2387-2400

Dretske, F. (1981), “Knowledge and the flow of information”, MA: MIT Press

Gradin, V.B., Kumar, P., Waiter, G., Ahearn, T., Stickle, C., Milders, M., Reid, I., Hall, J. & Steele, J.D. (2011), “Expected value and prediction error abnormalities in depression and schizophrenia”, Brain (134), pp.1751-1764

Hohwy, J. (2014), “The self-evidencing brain”, Nous, advance online access available at < >, retrieved 27/03/2015

Hohwy, J. & Rajan, V. (2012), “Delusions as forensically disturbing perceptual inferences”, Neuroethics (5), pp.5-11

Hohwy, J., Roepstorff, A. & Friston, K. (2008), “Predictive coding explains binocular rivalry: An epistemological review”, Cognition (108:3), pp.687-701

Jacob, P. (2014), “Intentionality”, in Zalta, E.N. (ed.) The Stanford Encyclopedia of Philosophy, available at, accessed 31/03/2015

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

Levy, N. (2014) “Addiction as a disorder of belief”, Biological Philosophy (29), pp.337-355

Prinz, J. (2006), “Beyond appearances: The content of sensation and perception”, in Gendler, T.S. & Hawthorne, J., Perceptual Experience, pp.434-459

Share your thoughts on this!

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s