The Bayesian Brain and Predictive Processing: A Critique, #2.

But you can also download and read or share a .pdf of the complete text of this essay, including the REFERENCES, by scrolling down to the bottom of this post and clicking on the Download tab.

3. Part B: Critique

3.1 Critical Challenges to Bayesian Brain and Predictive Processing Frameworks

While Bayesian brain and predictive processing frameworks have gained substantial followings and influence in cognitive neuroscience, they face significant theoretical, empirical, and philosophical challenges. We now examine fundamental critiques including: the computational intractability of Bayesian inference in realistic neural systems; lack of clear neural mechanisms for implementing probabilistic computations; unfalsifiability concerns stemming from post-hoc explanatory flexibility; alternative explanations for supposedly supportive evidence, neglect of action-first and embodied approaches; reductionism about consciousness and phenomenology, and conceptual confusion between descriptive and mechanistic claims. These challenges suggest that while Bayesian and predictive processing approaches offer valuable heuristics, their status as fundamental theories of brain function remains highly questionable.

The Limits of Theoretical Unification

The Bayesian brain hypothesis and predictive processing framework have achieved remarkable influence, offering seemingly unified accounts of perception, action, learning, attention, and psychopathology. However, this theoretical unification may come at the cost of empirical specificity and mechanistic clarity. As these frameworks have expanded to encompass increasingly diverse phenomena, critical questions have emerged about their explanatory power, falsifiability, and relationship to actual neural mechanisms.

We examine major challenges to Bayesian and predictive processing approaches, organized into computational, empirical, conceptual, and philosophical categories. Rather than dismissing these frameworks entirely, this critique aims to clarify their limitations and identify where claims exceed supporting evidence.

3.2 Computational Intractability and the Tractability Problem

The Curse of Dimensionality

A fundamental challenge for Bayesian brain theories concerns computational tractability. Exact Bayesian inference is computationally intractable for realistic problems involving high-dimensional state spaces and complex generative models (Kwisthout, Wareham, and van Rooij, 2011). The number of competing possible hypotheses grows exponentially with the dimensionality of the problem, creating what computer scientists call “the curse of dimensionality.”

Consider visual perception: estimating the three-dimensional structure of a scene from two-dimensional retinal images involves solving an inverse problem with astronomical numbers of possible interpretations. Even with hierarchical structure, the combinatorial explosion of possible hypotheses at multiple levels makes exact Bayesian inference computationally prohibitive (Yuille and Kersten, 2006; Bowers and Davis, 2012).

Proponents respond that the brain implements approximate Bayesian inference using sampling methods, variational approximations, or heuristics (Sanborn and Chater, 2016). However, this response weakens the theory’s explanatory power. If the brain uses approximations that deviate systematically from optimal Bayesian inference, then behavioral data showing suboptimal performance cannot distinguish between genuine Bayesian computation and alternative non-Bayesian mechanisms that produce similar outputs (Jones and Love, 2011).

The Problem of Priors

Bayesian inference requires specifying prior probability distributions over hypotheses. But where do these priors come from, and how are they represented neurally? The theory faces a dilemma: either priors are innate (raising evolutionary implausibility for highly specific distributions), or they are learned from experience (creating circularity, as learning itself requires priors) (Bowers and Davis, 2012).

Moreover, realistic predictive processing models require structured, often hierarchical priors that encode sophisticated knowledge about causal structure in the environment. How such complex probabilistic knowledge is acquired, represented, and updated neurally remains deeply unclear (Marcus and Davis, 2013). The specification of appropriate priors often requires substantial domain expertise, raising questions about whether unaided neural learning could discover such priors.

3.3 Neural Implementation Mysteries

Perhaps the most serious computational challenge concerns neural implementation. Despite decades of research, no clear neural mechanisms have been identified for representing and manipulating probability distributions (Rahnev and Denison, 2018). How do neurons encode probability distributions? How are Bayesian computations—multiplication of likelihoods and priors, normalization by marginal probabilities—implemented in neural circuits?

Probabilistic population codes have been proposed as implementation mechanisms (Ma, Beck, Latham, and Pouget, 2006), but these proposals face difficulties. Population codes require precise tuning of neural variability and correlation structures that may not exist in real neural populations (Beck, Ma, Pitkow, Latham, and Pouget, 2012). Alternative proposals invoke sampling mechanisms, but these face their own implementation challenges and timing problems (Orbán, Berkes, Fiser, and Lengyel, 2016).

3.4 Empirical Challenges and Alternative Explanations

Underdetermination and Flexibility

A pervasive criticism of Bayesian and predictive processing frameworks concerns their explanatory flexibility. Because these theories involve multiple free parameters (prior distributions, likelihood functions, precision weightings), they can be fit to almost any behavioral or neural data after the fact (Bowers and Davis, 2012; Jones and Love, 2011).

This flexibility undermines falsifiability. When predictions fail, defenders can always invoke different priors, alternative precision weightings, or approximations that deviate from optimal inference. The theory becomes unfalsifiable—capable of accommodating any outcome through parameter adjustment (Marcus and Davis, 2013).

Consider schizophrenia: predictive processing accounts explain hallucinations as excessive precision on internally generated predictions (Adams et al., 2013), but delusions as failures to update beliefs despite prediction errors. The theory thus explains both excessive and insufficient belief updating as manifestations of the same underlying framework, raising questions about what evidence could falsify these accounts.

Alternative Explanations for Key Evidence

Much evidence cited in support of Bayesian and predictive processing admits alternative explanations:

Cue Integration: While optimal cue integration appears to support Bayesian inference (Ernst and Banks, 2002), simple weighted averaging mechanisms without probabilistic representations can produce similar behavior (Landy, Banks, and Knill, 2011). The match to Bayesian optimality may reflect task-specific learning rather than general Bayesian principles.

Repetition Suppression: Neural adaptation to repeated stimuli, interpreted as reduced prediction error (Summerfield et al., 2008), can equally be explained by fatigue, habituation, or resource optimization without invoking predictions (Grotheer and Kovács, 2016). Single-cell recordings show repetition suppression even in early sensory areas where predictive coding accounts seem implausible.

Contextual Effects: Prior knowledge effects on perception, often cited as Bayesian priors, could reflect simpler associative mechanisms or learned affordances without probabilistic computation (Bowers and Davis, 2012). Associative learning can produce behavior superficially resembling Bayesian inference without implementing probabilistic calculations.

Failures of Bayesian Optimality

Extensive research documents systematic deviations from Bayesian optimality in human judgment and perception (Kahneman, 2011; Gigerenzer and Gaissmaier, 2011). People show base-rate neglect, conjunction fallacies, and numerous other systematic biases that violate Bayesian principles. While some deviations can be explained as rational responses to computational constraints, many biases appear robustly irrational even when computational costs are minimal.

Moreover, developmental research shows children often fail to integrate evidence in Bayesian-optimal ways, even in simple tasks (Téglás et al., 2011). If Bayesian inference is fundamental to brain function, why does it emerge slowly and incompletely through development?

3.5 Conceptual and Theoretical Problems

The Representation Problem

Predictive processing claims the brain represents hierarchical generative models, but what does this representation consist in? The theory remains vague about what neural states count as representing probability distributions versus merely correlating with environmental statistics (Gładziejewski, 2016).

This vagueness allows predictive processing to avoid empirical constraints. Any neural activity that varies with environmental statistics can be interpreted as representing a generative model, making the theory difficult to distinguish from claims that neural activity simply responds to statistical regularities (Williams, 2018).

The Dark Room Problem

If organisms minimize prediction error, why don’t they seek dark, unchanging rooms where sensory input is perfectly predictable (Friston, Thornton, and Clark, 2012)? Proponents respond that organisms have homeostatic setpoints requiring regular sensory inputs (hunger, thirst, etc.), but this response undermines claims that prediction error minimization is fundamental—it shows organisms pursue goals that sometimes increase prediction error.

Active inference attempts to resolve this by claiming organisms sample sensory data to reduce uncertainty about hidden states (Friston et al., 2012). However, this adds complexity and raises new questions about how organisms balance exploration (seeking informative prediction errors) with exploitation (confirming existing predictions).

The Explanatory Span Problem

As the predictive processing model has been extended to explain increasingly diverse phenomena, it risks becoming so general that it explains everything and therefore nothing (Klein, 2018). When a theory explains perception, action, attention, consciousness, emotion, learning, development, and psychiatric disorders through the same core mechanism, skepticism is warranted about whether genuine explanatory work is being done versus post-hoc redescription.

Different phenomena may require different explanatory frameworks rather than forced unification under a single principle. The drive for theoretical parsimony may obscure important mechanistic differences between perceptual inference, motor control, and high-level cognition.

3.6 Alternative Frameworks and Neglected Perspectives

Ecological and Action-First Approaches

Ecological psychology, in the tradition of J.J. Gibson, challenges the assumption that perception requires inference from impoverished sensory data (Gibson, 1979; Chemero, 2009). According to this view, ambient energy arrays contain rich information that specifies environmental properties directly, without requiring probabilistic inference.

Predictive processing assumes poverty of the stimulus—that sensory data is ambiguous and requires top-down disambiguation. But ecological approaches argue that organisms actively sample informative aspects of structured environments, reducing inferential demands (Noë, 2006). The theory may overestimate the brain’s inferential burden by underestimating environmental information structure.

Enactive and sensorimotor approaches similarly emphasize action and embodiment over internal representation (Varela et al., 1991; O’Regan and Noë, 2001). Rather than constructing internal models, organisms enact perceptual experience through skilled sensorimotor engagement with environments. These approaches question whether rich internal generative models are necessary or whether simpler sensorimotor contingencies suffice.

Simple Heuristics and Fast-and-Frugal Cognition

Research on ecological rationality demonstrates that simple heuristics often outperform complex Bayesian computations in realistic environments (Gigerenzer and Gaissmaier, 2011). Fast-and-frugal heuristics—simple decision rules that ignore information—frequently match or exceed Bayesian performance while requiring vastly less computation.

This suggests that apparent Bayesian optimality in behavior may reflect evolution selecting simple heuristics that perform well in natural environments, rather than implementing general Bayesian inference machinery (Todd and Gigerenzer, 2012). The brain may consist of many specialized systems using simple rules rather than a unified Bayesian inference engine.

Reinforcement Learning Alternatives

Standard reinforcement learning (RL) provides alternative accounts of learning and decision-making without requiring probabilistic inference over generative models (Sutton and Barto, 2018). While predictive processing proponents argue that RL is a special case of active inference (Friston et al., 2009), critics maintain that standard RL better captures actual neural mechanisms in dopaminergic systems and basal ganglia (Niv, 2009).

RL models make specific predictions about neural signals (reward prediction errors) that are empirically supported, whereas predictive processing’s claims about neural precision weighting and hierarchical prediction errors remain more speculative (Rescorla, 2016).

3.7 Philosophical and Phenomenological Critiques

The Phenomenological Inadequacy Problem

Phenomenological philosophers argue that predictive processing fundamentally mischaracterizes conscious experience (Ratcliffe, 2008; Gallagher and Allen, 2016). Conscious perception feels like direct openness to the world, not like unconscious inference from sensory data. The lived experience of perception includes a sense of presence and engagement that seems incompatible with treating perception as hypothesis testing.

Merleau-Ponty’s phenomenology emphasizes the primacy of perceptual engagement over cognitive representation (Merleau-Ponty, 1945/2012). Perception involves skilful bodily engagement with meaningful environments, not computational inference over abstract representations. Predictive processing might intellectualize perception, mistaking scientific models for lived experience.

The Hard Problem Remains

Despite claims that predictive processing illuminates consciousness (Hohwy, 2013; Clark, 2016), “the hard problem of consciousness”—why there is subjective experience at all—remains untouched (Chalmers, 1996). Explaining how prediction error minimization generates particular patterns of neural activity does not explain why these patterns should feel like anything.

Predictive processing might confuse explanations of cognitive function (access consciousness, reportability) with explanations of phenomenal consciousness (subjective experience). The theory addresses the former while leaving the latter mysterious, despite suggestions that it offers progress on consciousness (Seth, 2021).

Idealism and Anti-Realism Worries

Some critics argue that predictive processing implies problematic idealism or anti-realism about the external world (Bruineberg, Kiverstein, and Rietveld, 2018). If perception constructs reality from internal models rather than detecting mind-independent features, what grounds our confidence in realism?

While proponents argue that prediction error keeps internal models anchored to reality (Clark, 2016), critics maintain that this response doesn’t fully address the problem. If all we access are our own predictions, how do we know these predictions track a mind-independent world rather than merely achieving internal coherence?

3.8 Evaluating Neural Evidence

Ambiguous Neural Signals

Neural signals interpreted as prediction errors could alternatively reflect:

Neural signals interpreted as prediction errors could alternatively reflect:

Novelty detection: Responses to unexpected stimuli without predictive coding.

Attention effects: Enhanced processing of surprising events through different mechanisms.

Memory mismatch: Comparison with memory traces rather than predictions.

Adaptation: Habituation to repeated stimuli without prediction.

Single-cell and population recordings rarely provide sufficient detail to distinguish these alternatives (Heilbron and Chait, 2018). The interpretation of neural responses as prediction errors often reflects theoretical commitment rather than empirical necessity.

Missing Neural Evidence

Despite decades of research, several key predictions of predictive coding remain unsupported:

Separate error and prediction units: Clear anatomical separation of error and prediction neurons has not been consistently demonstrated (Keller and Mrsic-Flogel, 2018).

Precision modulation mechanisms: Neural mechanisms for implementing precision weighting remain unclear (Moran et al., 2013).

Hierarchical error propagation: Direct evidence for error signals propagating up cortical hierarchies is limited (Walsh et al., 2020).

The theory makes specific architectural predictions about cortical microcircuits that have not been confirmed. While some evidence is consistent with predictive coding, alternative architectures could produce similar macroscopic patterns.

3.9 Methodological Concerns

Model Comparison Problems

Studies claiming to support Bayesian models often fail to compare them against adequately-specified alternatives (Jones and Love, 2011). Non-Bayesian models could fit the same data equally well or better if given equivalent flexibility and fitting procedures.

Meta-analyses suggest that Bayesian models sometimes fit behavior through parameter flexibility rather than capturing genuine probabilistic reasoning (Eberhardt and Danks, 2011). More rigorous model comparison using techniques like cross-validation, out-of-sample prediction, and comparison against deliberately constructed alternative models is needed.

Publication Bias and Confirmatory Research

The field may suffer from publication bias favoring positive results supporting Bayesian and predictive processing frameworks (Ioannidis, 2005). Studies showing failures of Bayesian optimality or alternative explanations for supposedly supportive evidence may be underrepresented in the literature.

Additionally, much research is confirmatory rather than exploratory—designed to demonstrate Bayesian principles rather than rigorously test whether Bayesian models outperform alternatives. This confirmatory emphasis can create an illusion of stronger support than warranted.

3.10 Integration Challenges

Relationship to Neuroscience

The gap between high-level computational principles and detailed neural mechanisms remains wide (Carandini, 2012). Merely claiming that the brain “implements” Bayesian inference or predictive coding does not itself specify actual neural algorithms and circuits. The theory risks remaining at Marr’s computational level without successfully connecting to algorithmic and implementation levels.

Neuroscience has identified numerous specialized systems (sensory processing, motor control, memory, attention) with distinct neural mechanisms. Predictive processing’s claim that all these systems share a common computational principle may underestimate neural heterogeneity and specialization (Anderson, 2014).

Relationship to Evolution

Evolutionary considerations raise questions about whether unified Bayesian inference would evolve versus collections of specialized mechanisms (Barrett and Kurzban, 2006). Natural selection typically produces specialized adaptations rather than domain-general solutions. The claim that the brain implements general Bayesian principles across domains may be evolutionarily implausible.

Moreover, evolution operates through satisficing rather than optimizing—selecting “good enough” solutions rather than optimal ones. The brain’s mechanisms may reflect evolutionary tinkering more than principled optimization, contrary to predictive processing’s emphasis on minimizing free energy (Godfrey-Smith, 1996).

3.11 Constructive Paths Forward

Limited Domain Applications

Rather than universal theories of brain function, Bayesian and predictive processing approaches might be most useful as frameworks for understanding specific domains where inference over internal models is plausible—perhaps aspects of perception, certain types of learning, and some high-level cognitive processes (Rahnev and Denison, 2018).

Affirming a more limited scope for the Bayesian theory would increase empirical testability and allow more productive engagement with domain-specific neural mechanisms. The frameworks could serve as useful tools for specific applications without claiming to be fundamental theories of all brain function.

Integration with Alternative Approaches

Rather than asserting hegemony, predictive processing could be integrated with ecological, enactive, and reinforcement learning approaches (Bruineberg et al., 2018). Different brain systems may use different computational strategies, with some implementing prediction-based inference while others use simpler heuristics, reactive mechanisms, or action-based learning.

Pluralistic approaches that recognize mechanistic diversity may better capture neural reality than attempts to subsume everything under unified principles (Anderson, 2014).

Improved Empirical Rigor

Future research should employ:

Rigorous model comparison against well-specified alternatives.

Direct neural tests of specific mechanistic predictions.

Cross-validation and out-of-sample prediction to test generalization.

Adversarial collaboration between proponents and critics.

Transparency about model flexibility and post-hoc adjustments.

These methodological improvements would strengthen the empirical foundations and clarify the scope of valid applications.

Download

Against Professional Philosophy is a sub-project of the online mega-project Philosophy Without Borders, which is home-based on Patreon here.

Please consider becoming a patron!

Against Professional Philosophy

A Co-Authored Anarcho-Philosophical Diary

The Bayesian Brain and Predictive Processing: A Critique, #2.

Related

Share this:

Related