PP (Predictive processing; for this terminology, see Clark 2013a) turns a traditional picture of perception on its head. According to that once-standard picture (Marr 1982), perceptual processing is dominated by the forward flow of information transduced from various sensory receptors. As information flows forward, a progressively richer picture of the real-world scene is constructed. The process of construction would involve the use of stored knowledge of various kinds, and the forward flow of information was subject to modulation and nuancing by top-down (mostly attentional) effects. But the basic picture remained one in which perception was fundamentally a process of “bottom-up feature detection”. In Marr’s theory of vision, detected intensities (arising from surface discontinuities and other factors) gave way to detected features such as blobs, edges, bars, “zero-crossings”, and lines, which in turn gave way to detected surface orientations leading ultimately (though this step was always going to be problematic) to a three-dimensional model of the visual scene. Early perception is here seen as building towards a complex world model by a feedforward process of evidence accumulation. Traditional perceptual neuroscience followed suit, with visual cortex (the most-studied example) being “traditionally viewed as a hierarchy of neural feature detectors, with neural population responses being driven by bottom-up stimulus features” (Egner et al. 2010, p. 16601). This was a view of the perceiving brain as passive and stimulus-driven, taking energetic inputs from the senses and turning them into a coherent percept by a kind of step-wise build-up moving from the simplest features to the more complex: from simple intensities up to lines and edges and on to complex meaningful shapes, accumulating structure and complexity along the way in a kind of Lego-block fashion.
Such views may be contrasted with the increasingly active views that have been pursued over the past several decades of neuroscientific and computational research. These views (Ballard 1991; Churchland et al. 1994; Ballard et al. 1997) stress the active search for task-relevant information just-in-time for use. In addition, huge industries of work on intrinsic neural activity, the “resting state” and the “default mode” (for a review, see Raichle & Snyder 2007) have drawn our attention to the ceaseless buzz of neural activity that takes place even in the absence of ongoing task-specific stimulation, suggesting that much of the brain’s work and activity is in some way ongoing and endogenously generated.
Predictive processing plausibly represents the last and most radical step in this retreat from the passive, input-dominated view of the flow of neural processing. According to this emerging class of models, naturally intelligent systems (humans and other animals) do not passively await sensory stimulation. Instead, they are constantly active, trying to predict the streams of sensory stimulation before they arrive. Before an “input” arrives on the scene, these pro-active cognitive systems are already busy predicting its most probable shape and implications. Systems like this are already (and almost constantly) poised to act, and all they need to process are any sensed deviations from the predicted state. It is these calculated deviations from predicted states (known as prediction errors) that thus bear much of the information-processing burden, informing us of what is salient and newsworthy within the dense sensory barrage. The extensive use of top-down probabilistic prediction here provides an effective means of avoiding the kinds of “representational bottleneck” feared by early opponents (e.g., Brooks 1991) of representation-heavy—but feed-forward dominated—forms of processing. Instead, the downward flow of prediction now does most of the computational “heavy-lifting”, allowing moment-by-moment processing to focus only on the newsworthy departures signified by salient (that is, high-precision—see section 3) prediction errors. Such economy and preparedness is biologically attractive, and neatly sidesteps the many processing bottlenecks associated with more passive models of the flow of information.
Action itself (more on this shortly) then needs to be reconceived. Action is not so much a response to an input as a neat and efficient way of selecting the next “input”, and thereby driving a rolling cycle. These hyperactive systems are constantly predicting their own upcoming states, and actively moving so as to bring some of them into being. We thus act so as to bring forth the evolving streams of sensory information that keep us viable (keeping us fed, warm, and watered) and that serve our increasingly recondite ends. PP thus implements a comprehensive reversal of the traditional (bottom-up, forward-flowing) schema. The largest contributor to ongoing neural response, if PP is correct, is the ceaseless anticipatory buzz of downwards-flowing neural prediction that drives both perception and action. Incoming sensory information is just one further factor perturbing those restless pro-active seas. Within those seas, percepts and actions emerge via a recurrent cascade of sub-personal predictions forged (see below) from unconscious expectations spanning multiple spatial and temporal scales.
Conceptually, this implies a striking reversal, in that the driving sensory signal is really just providing corrective feedback on the emerging top-down predictions.[1] As ever-active prediction engines, these kinds of minds are not, fundamentally, in the business of solving puzzles given to them as inputs. Rather, they are in the business of keeping us one step ahead of the game, poised to act and actively eliciting the sensory flows that keep us viable and fulfilled. If this is on track, then just about every aspect of the passive forward-flowing model is false. We are not passive cognitive couch potatoes so much as proactive predictavores, forever trying to stay one step ahead of the incoming waves of sensory stimulation.