Clark has emphasized the way in which PP departs from the standard picture in perceptual psychology, and from David Marr’s (1982) model of visual processing in particular (pp. 1–5). According to the standard account, the flow of information is “bottom-up,” as perceptual systems construct increasingly sophisticated representations based on the information transduced at the periphery. According to PP, perception involves the active prediction of the upcoming sensory input, “top-down.” Deviation from what is predicted, known as the prediction error, propagates upwards through the hierarchy until it is explained away by the Bayesian generative model.
Now I would like to add that the standard picture in perceptual psychology has been widely regarded as complementary to the standard picture in the philosophy of perception (see Tye 2000, for example). One central question in the philosophy of perception is the following: what is the content of perceptual states? Or, what does perception represent? The standard answer, in tune with Marr’s approach, is that perceptual systems represent the external world, more or less as it really is. As Marr puts it, the purpose of vision is “to know what is where by looking” (1982). This way of thinking about perceptual content is almost a commonplace in the philosophical literature (Lewis 1980, p. 239; Fodor 1987, Ch. 4; Dretske 1995, Ch. 1). Kathleen Akins has described how the orthodox conception regards the senses as “servile” in that they report on the environmental stimulus “without fiction or embellishment” (1996, pp. 350–351).
Since PP overturns the reigning model in perceptual psychology, one might now ask whether it also overturns the reigning model in the philosophy of perception. Here are two initial reasons to think that it does. First, according to PP, there is always an active contribution from the organism, or at least from a part of the organism. Perceptual states are generated internally and spontaneously by the ongoing dynamics of the generative model. Those states are constrained by perceptual sampling of the world, not driven by input from the world. Perceptual states are driven by the endogenous activity of the predictive brain. The relevant causal history of these states begins, if you will, within the brain, rather than from the outside. Each organism’s generative model is unique in that it has been formed and continuously revised according to the particular trajectory of that organism’s cycle of action and perception. As Clark himself puts it, the forward flow of sensory information is always “relative to specific predictions” (p. 6). These considerations make it clear that there can be variation in perceptual content for identical environmental conditions. Perceivers with different histories will have different predictions (Madary 2013, pp. 342–345). The degree of variation is an open question, but it is reasonable to expect variation.
A second reason to think that PP motivates a richer conception of perceptual content is that perception, according to PP, is not simply in the service of informing the organism “what is where.” One main feature of PP is that perception and action work together in the service of minimizing prediction error. Clark explains that in “active inference […] the agent moves its sensors in ways that amount to actively seeking or generating the sensory consequences that they […] expect” (2013, p. 6, also see his discussion on page 16). If this is right, then perception does not serve the purpose of simply reporting on the state of the environment. Instead, perception is guided by expectation. While the received view of perceptual content answers the question of “what is out there?”, PP suggests that perceptual content answers the question of “is this what I expected and tested via active inference?” In a way, PP simplifies perceptual content by replacing the goal of representing the world with the single guiding principle of error minimization.
These two points suggest an understanding of perceptual content as something that is deeply informed by the specific history and embodiment of the organism. The content of perception is a complex interplay between particular organisms and their particular environments. At least on the face of it, this way of considering perception suggests new challenges and interesting new theoretical options for philosophers interested in describing perceptual content. For one thing, it suggests that propositional content as expressed using natural language (Searle 1983, p. 40) may be ill-suited for the task of describing perceptual content. Natural language does not typically include reports about prediction-error minimization, nor does it capture the fine-grained differences in perceptual content that will arise due to slight variations in the predictions made by different organisms. The traditional account of perceptual content, following Marr, does not include such differences, and is thus better disposed to expression using natural language.
These new challenges for understanding perceptual content may offer at the same time a general lesson for understanding all mental content in a naturalistic manner. Let me explain. One of the main goals in the philosophy of psychology has been to naturalize intentionality, to give an account of the content of mental states in terms of the natural sciences (in non-mentalistic terms). Well-known attempts include causal co-variation (Fodor 1987, Ch. 4) and teleosemantics (Millikan 1984, 2004). All attempts have met with compelling counterexamples.[1] Importantly, one implicit presupposition in the debate is that mental content should be conceived along the lines of the traditional view of perceptual content sketched above. That is, mental states are thought to be about bits of the objective world considered independently of the particular organism who possesses those mental states. To use a standard example, my belief that there is milk in the refrigerator is true if and only if there is milk in the refrigerator. This belief is about bits of the objective world: milk and the refrigerator in particular. Nothing else about my mind is deemed relevant for understanding the content of that belief. To use the familiar phrase, beliefs have a mind-to-world direction of fit (based on Anscombe 1957, §32).
If my reading of PP is right, and perceptual content turns out to be a matter of the complex interaction between particular organisms and their environments, then the comfortable pre-theoretical mind/world distinction might need revision.[2] Recall the discussion above, in which I claimed that, on the new PP-inspired understanding of perception the question is about whether sensory stimulation fulfils the expectations of particular organisms. All perceptual states are thereby colored, as it were, by the mental lives of the organisms having those states. Organisms are not interested in what the world is like. Organisms are interested in sustaining their integrity and physical existence; they are interested in what the world is like relative to their own particular sensorimotor trajectory through the world, a trajectory that is partly determined by their phenotype (Friston et al. 2006). This refashioning of the mind/world relationship is unorthodox, but it is hardly new. Similar ideas can be found in von Uexküll’s Umwelt (1934), Merleau-Ponty’s discussion of sensory stimuli (1962, p. 79), Millikan’s “pushmi-pullyu” representations (1995), Akins’ narcissistic sensory systems (1996), Clark’s earlier work (1997, Ch. 1), and in Metzinger’s ego tunnel (2009, pp. 8–9).
Now return to the problem of naturalizing intentionality. If we replace the notion of a purely world-directed mental state with a world-relative-to-the-organism-directed mental state, then naturalizing intentionality must somehow incorporate the relationship between the organism and its world. One way to pursue this project is to make it a matter of biology and physics. All living organisms keep themselves far from thermodynamic equilibrium by continuously exchanging matter and energy with their environment (Haynie 2008). Perhaps intentionality can be recast in terms of the organism’s ongoing struggle to maintain itself as a living entity. This line of thought is central to the enactivist “sense-making” of Maturana, Varela, and Thompson (Maturana & Varela 1980; Thompson 2007). Crucially, it is also a central feature of Friston’s version of PP. According to Friston, prediction error minimization is a kind of functional description for the physical process of the organism’s minimizing free energy in its effort to maintain itself far from thermodynamic equilibrium (2013). Naturalizing intentionality may be just a matter of physics (see Dixon et al. 2014 for an implementation of this strategy for problem-solving tasks).
Before moving on to the next section, I should add two qualifications. First, the idea of perceptual content being partly determined by the particular history of the perceiver should not be misunderstood as some kind of radical relativism with regard to perceptual content. Even if perceptual content is partly determined by the details of the organism, it is also partly determined by the world itself. As proponents of PP frequently claim, our generative models mirror the causal structure of the world (Hohwy 2013, Ch. 1). The point I am emphasizing here is that the causal structure of the world that is extracted is a structure relative to the embodiment (see Clark this collection, section 2.4)—and perceptual history—of the perceiver. The causal structure mirrored by a chimpanzee’s generative model is, in important ways, unlike the causal structure mirrored by that of a catfish.
The second qualification has to do with my remark that naturalizing intentionality may be just a matter of physics. Even if one allows that the approach I sketched shows promise, it is important to emphasize the explanatory gulf that remains. The intentionality-as-physics approach might succeed in explaining a bacterium’s intentional directedness towards a sugar gradient (Thompson 2007, p. 74–75), but it is far from clear how it would apply to my belief that P—say, for example, that California Chrome won the Kentucky Derby in 2014.
The main argument of this section has been that PP motivates an understanding of perceptual content that is always organism-relative. Clark’s version of PP, while not in conflict with this idea, has not addressed it explicitly, especially as it relates to the philosophy of perception. My goal here has been to do just that.