Surfing Uncertainty

1. The Brain is a Prediction Machine: Perception as Controlled Hallucination

Perceiving, imagining, understanding, and acting are now bundled together, emerging as different aspects and manifestations of the same underlying prediction-driven, uncertainty-sensitive, machinery.

Constant Guessing. Our brains are ceaseless prediction engines, constantly generating hypotheses about incoming sensory signals. This process isn't passive reception but an active attempt to "guess" what we're experiencing, using stored knowledge to anticipate the sensory barrage. This radical perspective suggests that perception is akin to "controlled hallucination," where our internal models construct reality.

Prediction Error as News. When our brain's predictions don't perfectly match the actual sensory input, a "prediction error" signal is generated. This error is the crucial "news" that propagates through the system, driving either the refinement of our internal models (learning) or the adjustment of our current guesses (perception). Instead of a bottom-up flow of raw data, the brain primarily processes these discrepancies.

Unified Cognitive Framework. This prediction-driven framework offers a powerful, unifying account for various cognitive functions. Perception, imagination, understanding, and even action are seen as different facets of the same core process: minimizing prediction error. This perspective fundamentally re-imagines how our minds engage with and make sense of the world.

2. Self-Supervised Learning Builds Generative Models from Sensory Data

The prediction task, as we shall see, is thus a kind of ‘bootstrap heaven’.

Learning by Predicting. The brain learns about the world by continuously trying to predict its own evolving sensory states. The environment itself acts as a constant, free, and ubiquitous teacher, providing the feedback needed to compare predictions with actual sensory outcomes. This self-supervised learning mechanism allows the brain to build increasingly sophisticated internal models.

Generative Models Defined. A "generative model" is a multi-layered probabilistic representation that captures the statistical structure of sensory inputs by inferring the interacting hidden causes that give rise to them. For instance, in vision, a generative model learns how lower-level visual patterns (like retinal stimulations) are produced by inferred distal causes (objects, scenes, movements). This model allows the brain to "self-generate" sensory data from the top-down.

Unearthing Deep Structure. This prediction-driven, multi-level learning process is remarkably powerful, enabling the brain to uncover complex, nested structures and regularities in the world. Examples include:

• Recognizing handwritten digits by learning to generate them.
• Inferring grammatical rules by predicting the next word in a sentence.
• Identifying abstract causal relationships from raw data.
  This "bootstrap heaven" allows the system to acquire knowledge about its environment without extensive pre-categorized training.

3. Attention is the Precision-Weighting of Prediction Error

Attention, thus construed, is a means of variably balancing the potent interactions between top-down and bottom-up influences by factoring in their so-called ‘precision’, where this is a measure of their estimated certainty or reliability (inverse variance, for the statistically savvy).

Modulating Signal Reliability. The brain doesn't treat all sensory information equally; it continuously estimates the reliability or "precision" of its prediction error signals. This precision acts like a volume knob, determining how much influence a particular error signal has on updating predictions or driving action. High precision means the signal is trustworthy and should be taken seriously.

Dynamic Balance of Influence. This precision-weighting mechanism allows the brain to flexibly balance top-down expectations with bottom-up sensory input. For example, in heavy fog, visual input is deemed unreliable (low precision), so prior knowledge plays a larger role. If a clear patch appears, that local signal's precision increases, allowing it to drive perception more strongly. This dynamic adjustment is central to adaptive behavior.

Unifying Attentional Effects. This framework unifies various attentional phenomena, from basic "pop-out" effects to complex, task-driven focus. Attention is not a separate mechanism but an integral part of the inferential cascade, enhancing selected prediction errors to sharpen responses and bias competition. This explains how we can selectively focus on specific features or locations, optimizing information processing for current goals.

4. Action is a Self-Fulfilling Prophecy of Proprioceptive Predictions

As strange as it sounds, when your own behaviour is involved, your predictions not only precede sensation, they determine sensation.

Active Inference in Motion. Action, like perception, is driven by prediction error minimization. Specifically, the brain predicts the proprioceptive (internal bodily sensations of position, force, and movement) consequences of a desired action. These predictions are not passive forecasts but active commands that the body strives to fulfill.

Proprioceptive Predictions as Motor Commands. When the brain predicts a sequence of proprioceptive states corresponding to a movement (e.g., reaching for a cup), a prediction error arises because the body is not yet in those states. The motor system then acts to cancel this error, moving the body to match the predicted sensations. Thus, the prediction itself becomes a self-fulfilling prophecy, directly causing the action.

Blurring Sensory-Motor Boundaries. This "active inference" perspective dissolves the traditional distinction between sensory and motor cortex. Both are fundamentally engaged in top-down prediction, though they predict different kinds of sensory input (e.g., visual vs. proprioceptive). This unified computational strategy simplifies motor control by integrating elements like cost functions into the generative model's expectations about trajectories.

5. Perception, Imagination, and Action Form a Unified Cognitive Package

From the simple seeds of a generative-model-based account of online perception, there thus emerges a striking (and strikingly familiar) cognitive form.

Co-Emergence of Mental Faculties. The very same generative models that enable rich, world-revealing perception also underpin imagination and dreaming. Since these models are capable of reconstructing sensory signals from the top-down, they can also generate "virtual" sensory experiences in the absence of external input. This suggests a deep functional duality between perception and imagination.

Mental Time Travel. This capacity for endogenous generation extends to "mental time travel," allowing us to reconstruct past events (memory) and simulate possible future scenarios (planning). The brain flexibly recombines details from past experiences into novel configurations, using the same predictive machinery to explore non-actual possibilities as it does to perceive the present.

Integrated Understanding. This "cognitive package deal" means that understanding is not a separate, abstract process but is intimately intertwined with perception, imagination, and action. Our grip on a structured, meaningful world emerges from the continuous, multi-level attempts to predict and engage with sensory signals, making these faculties co-dependent and mutually reinforcing.

6. Context-Sensitivity and Dynamic Neural Reconfiguration are Key to Flexible Response

The representational capacity and inherent function of any neuron, neuronal population, or cortical area is dynamic and context-sensitive [and] neuronal responses, in any given cortical area, can represent different things at different times.

Pervasive Contextual Influence. The hierarchical nature of predictive processing, combined with flexible precision-weighting, ensures maximal context-sensitivity. Higher-level predictions provide rich contextual information that sculpts the responsiveness of lower-level neurons. This means that the "meaning" or function of neural activity is not fixed but dynamically determined by the broader context.

Sculpting Effective Connectivity. Precision-weighting acts as a powerful "neural gating" mechanism, rapidly reconfiguring patterns of "effective connectivity" (causal influence between neural systems). By adjusting the gain on specific prediction error signals, the brain can dynamically form and dissolve "Transiently Assembled Local Neural Subsystems" (TALoNS) tailored to current tasks and contexts.

Understanding Others' Actions. This dynamic reconfiguration is crucial for complex social cognition, such as understanding the intentions behind others' actions. By modulating the precision of proprioceptive predictions, we can deploy our own action-generating models to interpret observed behaviors, integrating contextual cues (e.g., an operating theatre) to infer intentions (e.g., "to cure" vs. "to hurt").

7. The Brain Seeks Affordances, Not Just Passive Descriptions of Reality

The brain, they conclude, is ‘a fundamentally prospective organ that is designed to use information from the past and the present to generate predictions about the future’.

Beyond "Virtual Reality." While perception involves internal models, it's misleading to call it a "virtual reality." The brain's primary goal is not to create an action-neutral, perfectly accurate internal depiction of the world. Instead, it's to enable adaptive, world-engaging action and avoid surprising encounters. The "veil of transduction" is overcome by the very act of inference.

Perceiving Action Possibilities. Prediction-driven learning reveals "affordances"—the possibilities for action and intervention that the environment offers to a specific organism. Our perceptual "take" on the world is thus inherently pragmatic, constantly conditioned by our own action repertoire, needs, and opportunities. The world is "parsed for action."

Affordance Competition. The brain continuously computes multiple, probabilistically weighted affordances in parallel. These potential actions compete for control, with precision estimations biasing the system towards the most behaviorally relevant options. This "affordance competition" ensures that perception and action are deeply intertwined, leading to a world that is continuously "enacted" and shaped by our engagement.

8. The "Productively Lazy" Brain Optimizes for Efficiency and Adaptive Success

The most efficient strategy is simply the (active) inference that minimizes overall complexity costs.

Efficiency Over Perfection. The predictive brain is "productively lazy," prioritizing computationally cheap and effective strategies over exhaustive, optimal ones. It aims to "satisfice"—find solutions that are "good enough" given time and resource constraints. This involves balancing the accuracy of predictions with the complexity of the models used.

Exploiting Body and World. Many frugal strategies involve distributing the problem-solving load across the brain, the active body, and the environment. Examples include:

• Passive dynamics: Using the biomechanics of the body for efficient movement (e.g., walking).
• Ecological coupling: Using sensory input to directly coordinate with the environment (e.g., an outfielder catching a fly ball by canceling optical acceleration).
• Environmental scaffolding: Actively manipulating the world to simplify tasks (e.g., using an abacus).
  These strategies minimize internal computation by leveraging external resources.

Dynamic Strategy Switching. Predictive processing allows for flexible, context-dependent switching between model-rich (deliberative) and model-free (habitual/heuristic) strategies. Precision-weighting determines which strategy is most reliable and efficient in a given situation, enabling the brain to form temporary, soft-assembled circuits that span neural, bodily, and environmental resources.

9. Sociocultural Scaffolding Profoundly Transforms Human Prediction

By constructing a succession of designer environments, such as the human-built worlds of education, structured play, art, and science, we repeatedly restructure our own minds.

Human-Built Environments as Training Grounds. Human cognition is uniquely shaped by immersion in "designer environments" and "patterned sociocultural practices." These include language, tools, institutions, and educational systems. These external structures provide novel statistical regularities that train and transform our predictive brains, enabling the acquisition of generative models with unprecedented reach and depth.

Language as a Precision Manipulator. Language plays a crucial "supra-communicative" role, acting as a powerful tool for manipulating precision. Words and phrases can:

• Induce specific expectations, boosting perception (e.g., hearing "zebra" helps detect a masked image).
• Temporarily alter the tuning of neural populations.
• Provide "artificial contexts" that modify the influence of top-down information.
  This allows for targeted self-manipulation of our own uncertainty estimations, enhancing cognitive flexibility.

Collective Prediction and Cumulative Knowledge. Sociocultural practices foster "continuous reciprocal prediction" among agents, leading to shared understandings and joint action. Language enables ideas to be externalized, becoming public objects for critique and refinement across generations. This "incremental downstream epistemic engineering" allows human groups to explore intellectual trajectories far beyond what any individual brain could achieve alone.

10. Uncertainty Estimation Underpins Normal and Atypical Experience

The primary pathology here is quintessentially metacognitive in nature: in the sense that it rests on a belief (the warning light reports precise information) about a belief (the engine is overheating).

Metacognitive Core of Experience. The brain's continuous estimation of the reliability ("precision") of its own prediction error signals is a fundamental metacognitive ploy. This process is crucial for shaping normal perception, action, and emotion, determining how much confidence we place in sensory evidence versus our prior beliefs.

Pathologies of Precision. Disturbances in this delicate balance can lead to a wide range of atypical experiences:

• Schizophrenia: Overweighting of prediction error, leading to persistent "false errors" that drive the formation of bizarre, self-confirming delusions and hallucinations. Reduced sensory attenuation also contributes to misattributions of agency.
• Autism: Attenuated influence of prior knowledge ("hypo-priors"), causing sensory hypersensitivity and an overwhelming barrage of "newsworthy" information, leading to sensory overload and a preference for predictable routines.
• Functional Motor/Sensory Symptoms: Overweighting of prediction error at intermediate sensorimotor levels, leading to abnormal sensations or movements that are genuinely experienced but lack organic cause, often reinforced by folk beliefs.

Emotion and Placebo Effects. This framework also illuminates emotion, seen as multi-level inferences combining interoceptive (internal bodily) and exteroceptive signals with predictions. Placebo effects, similarly, arise from top-down predictions of pain relief or improved function, modulated by their estimated precision, demonstrating the profound impact of expectation on subjective experience and physiological response.

11. We Actively Enact Our Worlds Through Continuous Prediction-Action Cycles

The organism and environment [are] bound together in reciprocal specification and selection.

Mutual Specification of Organism and Environment. Human agents do not passively perceive a pre-given world; they actively "enact" it through continuous prediction-action cycles. This "structural coupling" means the organism and its environment are bound in reciprocal specification, where the organism's actions shape the sensory inputs it receives, which in turn refine its internal models.

Active Sampling and World-Building. We selectively move our bodies and sense organs to "serve up" predicted patterns of stimulation, especially those promising high-reliability, task-relevant information. This active sampling continuously confirms and constructs our individual and species-specific "worlds." At larger scales, humans deliberately structure their physical and social environments (e.g., building cities, creating cultural practices) to generate new, predictable patterns that further train and transform their predictive brains.

A Meaningful, Action-Oriented Reality. The predictive brain, deeply intertwined with the body and a self-constructed world, delivers a meaningful, action-oriented grip on reality. This reality is continuously disclosed and shaped by our activity, needs, and sociocultural context. The fit between mind and world is not one of passive description but of active engagement, where prediction error minimization drives a dynamic, adaptive dance between inner expectations and external reality.

Last updated: December 29, 2025