1 Introduction
An increasingly popular theory in cognitive science claims that brains are essentially prediction machines (Hohwy 2013). The theory is variously known as the Bayesian brain (Knill & Pouget 2004; Pouget et al. 2013), predictive processing (Clark 2013; Clark this collection), and the predictive mind (Hohwy 2013; Hohwy this collection), among others; here we use the term PP (predictive processing). (See Table 1 for a glossary of technical terms.) At its most fundamental, PP says that perception is the result of the brain inferring the most likely causes of its sensory inputs by minimizing the difference between actual sensory signals and the signals expected on the basis of continuously updated predictive models. Arguably, PP provides the most complete framework to date for explaining perception, cognition, and action in terms of fundamental theoretical principles and neurocognitive architectures. In this paper I describe a version of PP that is distinguished by (i) an emphasis on predictive modelling of internal physiological states and (ii) engagement with alternative frameworks under the banner of “enactive” and “embodied” cognitive science (Varela et al. 1993).
I first identify an unusual starting point for PP, not in Helmholtzian perception-as-inference, but in the mid 20th-century cybernetic theories associated with W. Ross Ashby (1952, 1956; Conant & Ashby 1970). Linking these origins to their modern expression in Karl Friston’s “free energy principle” (2010), perception emerges as a consequence of a more fundamental imperative towards homeostasis and control, and not as a process designed to furnish a detailed inner “world model” suitable for cognition and action planning. The ensuing view of PP, while still fluently accounting for (exteroceptive) perception, turns out to be more naturally applicable to the predictive perception of internal bodily states, instantiating a process of interoceptive inference (Seth 2013; Seth et al. 2011). This concept provides a natural way of thinking of the neural substrates of emotional and mood experiences, and also describes a common mechanism by which interoceptive and exteroceptive signals can be integrated to provide a unified experience of body ownership and conscious selfhood (Blanke & Metzinger 2009; Limanowski & Blankenburg 2013).
Table 1: A glossary of technical terms.
The focus on embodiment leads to distinct interpretations of active inference, which in general refers to the selective sampling of sensory signals so as to improve perceptual predictions. The simplest interpretation of active inference is the changing of sensory data (via selective sampling) to conform to current predictions (Friston et al. 2010). However, by analogy with hypothesis testing in science, active inference can also involve seeking evidence that goes against current predictions, or that disambiguates multiple competing hypotheses. A nice example of the latter comes from self-modelling in evolutionary robotics, where multiple competing self-models are used to specify actions that are most likely to provide disambiguatory sensory evidence (Bongard et al. 2006). I will spend more time on this example later. Crucially, these different senses of active inference rest on the capacity of predictive models to encode counterfactual relations linking potential (but not necessarily executed) actions to their expected sensory consequences (Friston et al. 2012; Seth 2014b). It also implies the involvement of model comparison and selection—not just the optimization of parameters assuming a single model. These points represent significant developments in the basic infrastructure of PP.
The notion of counterfactual predictions connects PP with what at first glance seems to be its natural opponent: “enactive” theories of perception and cognition that explicitly reject internal models or representations (Clark this collection; Hutto & Myin 2013; Thompson & Varela 2001). Central to the enactive approach are notions of “sensorimotor contingencies” and their “mastery” (O'Regan & Noë 2001), where a sensorimotor contingency refers to a rule governing how sensory signals change in response to action. On this approach, the perceptual experience of (for example) redness is given by an implicit knowledge (mastery) of the way red things behave given certain patterns of sensorimotor activity. This mastery of sensorimotor contingencies is also said to underpin perceptual presence: the sense of subjective reality of the contents of perception (Noë 2006). From the perspective of PP, mastery of a sensorimotor contingency corresponds to the learning of a counterfactually-equipped predictive model connecting potential actions to expected sensory consequences. The resulting theory of PPSMC (Predictive Perception of SensoriMotor Contingencies), Seth 2014b) provides a much needed reconciliation of enactive and predictive theories of perception and action. It also provides a solution to the challenge of perceptual presence within the setting of PP: perceptual presence obtains when the underlying predictive models are counterfactually rich, in the sense of encoding a rich repertoire of potential (but not necessarily executed) sensorimotor relations. This approach also helps explain instances where perceptual presence seems to be lacking, such as in synaesthesia.
This is both a conceptual and theoretical paper. Space limitations preclude any significant treatment of the relevant experimental literature. However, even an exhaustive treatment would reveal that this literature so far provides only circumstantial support for the basics of PP, let alone for the extensions described here. Yet an advantage of PP theories is that they are grounded in concrete computational processes and neurocognitive architectures, giving us confidence that informative experimental tests can be devised. Implementing such an experimental agenda stands as a critical challenge for the future.