Brains, it has recently been argued, are essentially prediction machines. They are bundles of cells that support perception and action by constantly attempting to match incoming sensory inputs with top-down expectations or predictions. This is achieved using a hierarchical generative model that aims to minimize prediction error within a bidirectional cascade of cortical processing. Such accounts offer a unifying model of perception and action, illuminate the functional role of attention, and may neatly capture the special contribution of cortical processing to adaptive success. This target article critically examines this “hierarchical prediction machine” approach, concluding that it offers the best clue yet to the shape of a unified science of mind and action. Sections 1 and 2 lay out the key elements and implications of the approach. Section 3 explores a variety of pitfalls and challenges, spanning the evidential, the methodological, and the more properly conceptual. The paper ends (sections 4 and 5) by asking how such approaches might impact our more general vision of mind, experience, and agency.

Clark appears to be moving toward epistemic internalism, which he once rightly rejected. This results from a double over-interpretation of predictive coding's significance. First, Clark argues that predictive coding offers a Grand Unified Theory (GUT) of brain function. Second, he over-reads its epistemic import, perhaps even conflating causal and epistemic mediators. We argue instead for a plurality of neurofunctional principles.

Clark advertises the predictive coding (PC) framework as applying to a wide range of phenomena, including attention. We argue that for many attentional phenomena, the predictive coding picture either makes false predictions, or else it offers no distinctive explanation of those phenomena, thereby reducing its explanatory power.

A cornerstone of the target article is that, in a predictive coding framework, attention can be modelled by weighting prediction error with a measure of precision. We argue that this is not a complete explanation, especially in the light of ERP (event-related potentials) data showing large evoked responses for frequently presented target stimuli, which thus are predicted.

The sensory cortex has been interpreted as coding information rather than stimulus properties since Sokolov in 1960 showed increased response to an unexpected stimulus decrement. The motor cortex is also organized around expectation, coding the goal of an act rather than a set of muscle movements. Expectation drives not only immediate responses but also the very structure of the cortex, as demonstrated by development of receptive fields that mirror the structure of the visual world.

Clark examines the notion of the “predictive brain” as a unifying model for cognitive neuroscience, from the level of basic neural processes to sensorimotor control. Although we are in general agreement with this notion, we feel that there are many details that still need to be fleshed out from the standpoint of perception and action.

Clark's essay lays the foundation for a Bayesian account of the “projection” of consciously perceived properties: The expectations that our brains test against inputs concern the particular affordances that evolution has designed us to care about, including especially expectations of our own expectations.

Clark makes a convincing case for the merits of conceptualizing brains as hierarchical prediction machines. This perspective has the potential to provide an elegant and powerful general theory of brain function, but it will ultimately stand or fall with evidence from basic neuroscience research. Here, we characterize the status quo of that evidence and highlight important avenues for future investigations.

We propose, following Clark, that generative models also play a central role in the perception and interpretation of linguistic signals. The data explanation approach provides a rationale for the role of prediction in language processing and unifies a number of phenomena, including multiple-cue integration, adaptation effects, and cortical responses to violations of linguistic expectations.

Why do brains have so many connections? The principles exposed by Andy Clark provide answers to questions like this by appealing to the notion that brains distil causal regularities in the sensorium and embody them in models of their world. For example, connections embody the fact that causes have particular consequences. This commentary considers the imperatives for this form of embodiment.

In important ways, Clark's “hierarchical prediction machine” (HPM) approach parallels the research agenda we have been pursuing. Nevertheless, we remain unconvinced that the HPM offers the best clue yet to the shape of a unified science of mind and action. The apparent convergence of research interests is offset by a profound divergence of theoretical starting points and ideal goals.

A radical interpretation of the predictive coding approach suggests that the mind is “seamless” – that is, that cancellation of error signals can propagate smoothly from highest to lowest levels of the control hierarchy, dissolving a distinction between belief and perception. Delusions of alien control provide a test case. Close examination suggests that while they are evidence of predictive coding within the cortex, they are not evidence for the seamless interpretation.

A quantitative theory of reproductive decisions (Gowaty & Hubbell 2009) says that individuals use updated priors from constantly changing demographic circumstances to predict their futures to adjust actions flexibly and adaptively. Our ecological/evolutionary models of ultimate causes seem consistent with Clark's ideas and thus suggest an opportunity for a unified proximate and ultimate theory of Bayesian animal brains, senses, and actions.

We suggest that the hierarchical predictive processing account detailed by Clark can be usefully integrated with narrative psychology by situating personal narratives at the top of an individual's knowledge hierarchy. Narrative representations function as high-level generative models that direct our attention and structure our expectations about unfolding events. Implications for integrating scientific and humanistic views of human experience are discussed.

The target article focuses on the predictive coding of “what” and “where” something happened and the “where” and “what” response to make. We extend that scope by addressing the “when” aspect of perception and action. Successful interaction with the environment requires predictions of everything from millisecond-accurate motor timing to far future events. The hierarchical framework seems appropriate for timing.

The brain is involved in theory-laden cognitive processes. But there are two different theory-laden processes. In cases where the theory is based on facts, more facts can either falsify or confirm a theory. In cases where the theory is about the choice of a benchmark or a standard, more facts can only make a theory either more or less warranted.

We argue that brains generate predictions only within the constraints of the action repertoire. This makes the computational complexity tractable and fosters a step-by-step parallel development of sensory and motor systems. Hence, it is more of a benefit than a literal constraint and may serve as a universal normative principle to understand sensorimotor coupling and interactions with the world.

A behavioral drive directed solely at minimizing prediction error would cause an agent to seek out states of unchanging, and thus easily predictable, sensory inputs (such as a dark room). The default to an evolutionarily encoded prior to avoid such untenable behaviors is unsatisfying. We suggest an alternate information theoretic interpretation to address this dilemma.

Clark offers a powerful description of the brain as a prediction machine, which offers progress on two distinct levels. First, on an abstract conceptual level, it provides a unifying framework for perception, action, and cognition (including subdivisions such as attention, expectation, and imagination). Second, hierarchical prediction offers progress on a concrete descriptive level for testing and constraining conceptual elements and mechanisms of predictive coding models (estimation of predictions, prediction errors, and internal models).

Clark acknowledges but resists the indirect mind–world relation inherent in prediction error minimization (PEM). But directness should also be resisted. This creates a puzzle, which calls for reconceptualization of the relation. We suggest that a causal conception captures both aspects. With this conception, aspects of situated cognition, social interaction and culture can be understood as emerging through precision optimization.