Representing the Self in Predictive Processing

22 October 2018 | Elmarie Venter (Ruhr-Universität Bochum)

Who do you think you are? Or, less confrontationally, what ingredients (e.g. memories, beliefs, desires) go into the model of your self? In this post, I explore different conceptions of how the self is represented in the predictive processing (PP) framework.

At the core of PP is the notion that the brain is in the business of making predictions about the world, and that the brain is primarily an organ that functions to minimize prediction error (i.e. the difference between predictions about the state of the world and the observed state of the world) (Clark, 2017, p.727). Predictive processing necessitates modelling the causes of our sensory perturbations and since agents themselves are also such causes, a self-model is required under PP. 

The internal models of the self will include “…representations of the agent’s own body and its trajectories and interactions with other causes in the world” (Hohwy & Michael, 2017, p.367).

In this post I will discuss accounts of how the self is modelled under two PP camps: Conservative PP and Radical PP. Broadly speaking, Conservative PP holds that the mind is inferentially secluded from the environment — the body also forms part of the external environment. All prediction error minimization occurs behind an ‘evidentiary boundary’ which implies that the brain reconstructs the state of the world (Hohwhy, 2016, p.259). 

In contrast, Radical PP holds that representations of the world are a matter of embodied and embedded cognition (Dolega, 2017, p.6). Perceiving my self, other agents, and the world, is not a process of reconstruction but rather a coupled process between perception and action. How does the view of a self-model align with these versions of predictive processing? 

I will argue that Radical PP’s account of self-modelling is preferable because it avoids two key concerns that arise from Conservative PP’s modeling of the self.

On the side of Conservative PP, Hohwy & Michael (2017) conceive of the self-model as one that captures “…representations of the agent’s own body…” as well as hidden, endogenous causes, such as “…character traits, biases, reaction patterns, affections, standing beliefs, desires, intentions, base-level internal states, and so on” (Hohwy & Michael, 2017, p.369). On this view, the self is just another set of causes that is modelled in order to minimize prediction error. 

This view likens the model of the self to models of the environment and other people (and their mental states), and is in line with the Conservative PP account advocated by Hohwy (2016) under which there is an ‘evidentiary boundary’ between mind and world, behind which prediction error minimization takes place. Any parts of our body “…that are not functionally sensory organs are beyond the boundary… [and are] just the kinds of states that should be modelled in internal, hierarchical models of a (prediction error minimization) system.” (Hohwy, 2016, p.269).

As I see it, Conservative PP’s self-modelling (as described by Hohwy & Michael (2017)) is problematic in two ways:

In dealing with these concerns I propose that the self is represented along two dimensions: as-subject and as-object (a distinction that can be traced back to Wittgenstein’s Blue Book, and which can be fleshed out by appeal to debates on reference and intentionality). The fundamental idea here is that there is a certain kind of error — in identifying the person that something is true of (e.g. a bodily position or a mental state) — that can occur when identifying the self as-object which cannot occur in identifying the self as-subject (Longuenesse, 2017, p.20; Evans 1982). 

Imagine that I perceive a coffee mug in front of me, and once I have seen it I reach out my hand to grasp the mug in order to drink from it. Now envision a similar situation, in which I am acting like this while at the same time looking at myself in a mirror. 

In the latter situation I have two sources of information for obtaining knowledge about myself grasping the cup of coffee. One source of information is proprioceptive and kinaesthetic, and therefore provides me with information about myself from the inside. 

The other source of information is visual, and provides me with information from the outside. The latter source could provide me with information about the actions of other agents as well, whereas the former can only be a source of information about my own self.

Since I am represented in the content of my visual experience in the mirror scenario, I can misrepresent myself as the intentional object of that very visual experience. I could be mistaken with respect to whom I am seeing in the mirror grasping the coffee mug; I may mistakenly believe that I am in fact observing someone else grasping the cup. 

No such mistake is possible in the contrast case, in which I gain information about grasping the mug from a proprioceptive and kinaesthetic source. A more radical example of this distinction between self as-object and as-subject comes from individuals with somatoparaphrenia. Such individuals do not identify some parts of their body as their own, e.g. they may believe that their arm belongs to someone else, but they are not mistaken about who is identifying their arm as belonging to someone else (Kang, 2016; Vallar & Ronchi, 2009). 

Recanati (2007, pp.147–148) spells out this difference by distinguishing between the content and mode of an intentional state: “The content is a relativized proposition, true at a person, and the internal mode determines the person relative to which that relativized content is evaluated: myself”. With this distinction in mind, the problems with Conservative PP becomes clear: the agent and their body are not represented in the same way as any other distal state in the world. Instead of the agent and their body only forming part of the content of an intentional state (as Hohwy & Michael’s account would imply), they enter the state through the mode of perception as well.

Clark (2017, p.729) provides an analogy that illustrates the first problem with self-modelling under Conservative PP: “The predicting brain seems to be in somewhat the same predicament as the imprisoned agents in Plato’s “allegory of the cave”.” That is, under Conservative PP, distal states can only be inferred by the secluded brain, just as the prisoners in the cave can only infer what the shadows on the walls are shadows of. 

The consequence of this is that we have no direct (and, therefore, error-immune) access to our own bodies. However, as has been illustrated above, the self enters intentional states through mode (perceiving, imagining, remembering, etc.) as well as content, and this provides us with certain information that is immune from error. 

In contrast, Radical PP does not conceive of the body as a distal object. Instead, the agent’s body plays an active role in determining the sensory information that we have access to; it plays a fundamental role in how we sample, and act in, the world. This active role is such that certain information is available to us error free – even if I am mistaken about another agent grasping the cup, I cannot be mistaken that it is me that is seeing someone grasp the cup.

In this sense, Radical PP provides us with a preferable story about how whole embodied agents are models of the environment and minimize prediction error through a variety of adaptive strategies (Clark, 2017,  p.742).

The two dimensions of self can also shed light on the second concern with Conservative PP because this distinction illustrates how we perceive and interact with other agents. As discussed above, the self as-object enters intentional states as part of the content, and the self as-subject enters such states through mode

The world, including other agents and their mental states, only ever form part of the content of our intentional states. Referring back to the example spelled out above: another agent can only ever play the same role in perception as I do in the mirror case, i.e. as content of the intentional structure. I do not have access to other agents “from the inside,” however. 

For instance, I do not have the same access to the reasons behind others’ actions (are they grasping the cup to drink from it, to clear it from the table, to see if there is still coffee in it?), nor do I have access to whether the other agent will successfully grasp the mug (is their grip wide enough, do they have enough strength in their wrist?). There is thus a dimension of the self to which one has privileged access. We only have access to other agents through perceptual inference (i.e. by observing their behaviour and inferring its causes), whereas we have both perceptual and active inferential access to our own behaviours. 

Though Conservative PP proponents maintain that the secluded brain only has perceptual inferential access to our own body (Hohwy, 2016, p.276), there is something markedly different in what enables us to model the causes of our own behaviour and mental states to that of other agents. I have proprioceptive, kinaesthetic, and interoceptive access to information about myself; I only have exteroceptive information about other agents.

For Conservative PP, the body (and by extension, the self) is just another object in the world that receives commands to act in service of prediction error minimization. I have highlighted two concerns about this view: the body is treated as a distal object, and the body (and self) placed on the same side of the evidentiary boundary as other agents. This means that the dimension of self which is immune to error through misidentification is not accommodated, and the marked difference in our access to information about our own states and those of other agents is ignored. 

Radical PP, however, avoids both concerns by taking into account the two representational dimensions of the self and employing an embodied approach to cognition. The Radical PP account therefore provides a more refined version of self-modeling. 

My beliefs, desires, and bodily shape can all be inferred in the model of self-as-object, but self-as-subject captures the part of the self that is not inferred: it contains information about me and my body from the inside, which is an essential part of who we think we are.


Clark, A., 2017. Busting Out: Predictive Brains, Embodied Minds, and The Puzzle of The Evidentiary Veil. Noûs, 51(4): 727–753.

Dolega, K., 2017. Moderate Predictive Processing. In T. Metzinger & W. Wiese (Eds.) Philosophy and Predictive Processing. Frankfurt Am Main: MIND Group.

Evans, J., 1982. The Varieties of Reference. Oxford: Clarendon Press.

Friston, K. J. and Stephan, K. E., 2007. Free-energy and the Brain. Synthese, 159(3): 417–458.

Hohwy, J., 2016. The Self-Evidencing Brain. Noûs, 50(2): 259–285.

Hohwy, J. and Michael, J., 2017. Why Should Any Body Have A Self? In F. de Vignemont & A. Alsmith (Eds.) The Subject’s Matter: Self-Consciousness And The Body. Cambridge, Massachusetts: MIT Press.

Kang, S. P., 2016. Somatoparaphrenia, the Body Swap Illusion, and Immunity to Error through Misidentification. The Journal of Philosophy, 113(9): 463–471.

Longuenesse, B., 2017. I, Me, Mine: Back To Kant, And Back Again. New York: Oxford University Press.

Michael, J. and De Bruin, L., 2015. How Direct is Social Perception. Consciousness and Cognition, 36: 373–375.

Recanati, F., 2007. Perspectival Thought: A Plea For (Moderate) Relativism. Clarendon Press.

Thompson, E. and Varela, F. J., 2001. Radical Embodiment: Neural Dynamics and Consciousness. Trends in Cognitive Sciences, 5(10): 418–425.

Vallar, G. and Ronchi, R., 2009. Somatoparaphrenia: A Body Delusion. A Review of the Neuropsychological Literature. Experimental Brain Research, 192(3): 533–551.

Wittgenstein, L. 1960. Blue Book. Oxford: Blackwell.