Abstract
Andy Clark’s book presents the mind as an active prediction-generating system rather than a passive receiver of sensory information. Clark argues that perception results from the brain’s continual interaction between prior expectations and incoming sensory signals.
Context
In Surfing Uncertainty: Prediction, Action, and the Embodied Mind (2019) Andy Clark combines traditions that are often treated separately from Kantian prediction to computational Bayesianism. His project is not only a theory of the brain, it is an attempt to naturalise older philosophical ideas about perception, agency, and situated experience
Kant influenced Andy Clark mainly through the idea that perception is active rather than passive: the mind organises sensory information instead of simply recording the world. Clark naturalises this insight through predictive processing, arguing that the brain constantly generates and revises predictions about sensory input. He also extends Kant’s emphasis on the perceiving subject by arguing that cognition depends on the body, environment, and external tools, not just the brain.
The key difference is that Kant gives a transcendental philosophical account, while Clark offers an empirical, scientific one. Clark’s predictive-processing model can be read as a naturalised, neuroscientific development of Kantian themes.
Hermann von Helmholtz influenced Andy Clark through the idea of unconscious inference: perception is not passive reception but the brain’s interpretation of sensory signals using prior expectations. Clark develops this idea in his theory of predictive processing, where the brain continually predicts incoming information and updates its models when errors occur.
Helmholtz provided the early psychological and physiological insight. Clark connects it to computational neuroscience and extends it to embodied and active cognition.
William James influenced Andy Clark mainly through pragmatism and functionalism. James viewed the mind as an active process shaped by bodily needs and interaction with the environment, rather than as a detached inner substance. Perception matters because it helps an organism cope with its environment. Clark develops similar ideas in his theories of embodied, embedded, and extended cognition. Like James, he emphasises that thinking is connected to action, perception, habit, and practical problem-solving. Clark’s work also extends James’s functionalist approach by arguing that cognitive processes can include the body, environment, and external tools.
James's influence is indirect: he supplied a pragmatist vision of mind as dynamic, action-oriented, and environmentally situated, which aligns closely with Clark’s embodied and extended approach.
John Dewey’s pragmatism supports Clark’s rejection of a sharp division between perception, thought, and action. Cognition is part of an organism’s ongoing activity in the world, rather than an isolated inner process. Clark’s work in embodied, situated, and extended cognition develops several ideas that are strongly Deweyan:
Dewey rejected the idea that cognition is something occurring solely inside the head. Clark similarly treats cognition as an ongoing activity of an organism engaged with its environment. For Dewey, perception and thought arise through the continuous interaction between an organism and its surroundings. Clark’s theory of extended cognition makes a related claim: tools, bodily actions, and environmental structures can become functionally integrated into cognitive processes.
Pragmatism and problem-solving. Dewey conceived thinking as a practical response to problematic situations: organisms use habits, tools, and experimentation to reorganize their circumstances. Clark’s account of cognition similarly emphasises action-oriented intelligence. Cognitive systems are designed to help agents cope, predict, and solve problems in real environments.
The relationship should not be overstated. Clark’s central theoretical influences also include connectionism, neuroscience, artificial intelligence, cybernetics, and phenomenology. He does not simply reproduce Dewey’s philosophy. Rather, he naturalises and scientifically reformulates Deweyan themes: cognition as embodied action, intelligence as environmentally situated, and tools as potentially constitutive parts of thinking. Scholarship explicitly connects Dewey’s pragmatism with extended-cognition theory, including Clark’s work.
Maurice Merleau-Ponty influenced Clark particularly in the development of the embodied and extended mind theses. Clark draws on Merleau-Ponty’s phenomenology to challenge the traditional "computational" view of the mind, which treats the brain as a central processor operating on internal representations of an external world. Perception is rooted in bodily capacities and practical involvement with the world, not merely in the construction of internal pictures.
The primary influence lies in Merleau-Ponty’s concept of the body-schema. Merleau-Ponty argued that the body is not merely a physical object that the mind controls, but the very medium through which we have a world. He described the body-schema as an implicit, pre-reflective awareness of our physical capabilities and our position in space. Clark translates this phenomenological insight into a cognitive framework, suggesting that our intelligence is not "trapped" in the skull but is distributed across our bodily movements and the tools we use.
This influence is most evident in Clark's theories on tool use and integration. Merleau-Ponty famously described how a blind man's cane ceases to be an external object and becomes an extension of his sense of touch, the "perceiving" happens at the tip of the cane, not in the hand. Clark adapts this to explain how humans incorporate external technology into their cognitive architecture. Whether it is a hammer, a smartphone, or a prosthetic limb, Clark argues that the brain can "transparently" integrate these tools into its functional loop, effectively extending the boundaries of the mind.
Clark also utilises Merleau-Ponty's rejection of the strict subject-object duality. By emphasising the sensorimotor coupling between an agent and its environment, Clark moves away from the idea that the mind creates a "map" of the world. Instead, he posits that cognition is an active process of exploration and interaction, mirroring Merleau-Ponty's view that perception is an active engagement with the environment rather than a passive reception of data.
Bayesian influence. (Bayesian probability is the name given to several related interpretations of probability as a measure of confidence in our knowledge – the strength of beliefs, hypotheses etc. – rather than a frequency.)
In his more recent work, Clark has integrated the merleau-pontian phenomenological insights with Bayesianism to create a comprehensive model of Predictive Processing. He posits that the brain functions as a Bayesian prediction engine, constantly generating top-down hypotheses about the world and using sensory input to correct them. This framework transforms the phenomenological "lived experience" into a computational process of minimising prediction error through a combination of perception and action.
Ultimately, Bayesianism provides the "how" to phenomenology's "what." While phenomenology describes the feeling of being an embodied agent interacting with an environment, Bayesian active inference explains the mechanism, showing how the brain manages uncertainty, directs attention through precision weighting, and drives movement to align reality with its internal expectations. Together, these influences allow Clark to argue that the mind is an active, embodied, and extended system that "surfs" the uncertainty of the world.
Clark’s own extended-mind theory is explained in his book The Extended Mind, coauthored with Chalmers. The book develops his earlier argument that cognition can extend beyond the brain into tools, language, social practices, and environmental structures. Prediction is not only something the brain does internally; bodily and environmental resources can help organise cognitive activity.
The evolution of Clark's thinking through his texts
In Being There (1997), Clark argues that thinking depends on the brain, body, and environment working together. Intelligence is not produced by the brain alone.
In The Extended Mind (1998), written with David Chalmers, Clark argues that external tools, such as notebooks and calculators, can become part of cognition when they function like memory or reasoning. Natural-Born Cyborgs (2003) and Supersizing the Mind (2008), expand this theory, claiming that humans naturally incorporate technology into their mental processes.
Whatever Next? (2013) and Surfing Uncertainty (2016), connect the extended mind to predictive processing. He argues that the brain constantly predicts the world, while the body and environment help test and update those predictions.
Overall, Clark moves from arguing that cognition is embodied to claiming that it can extend into tools, technology, and the surrounding world.
Download Surfing Uncertainty
Summary
Preface: Meat That Predicts
In the preface to Surfing Uncertainty, Andy Clark addresses the fundamental philosophical mystery of how physical matter — referred to colloquially and provocatively as "meat" — can give rise to complex mental states like consciousness, dreaming, and intelligent action.
He introduces the concept of prediction as a primary clue to solving this mystery. Clark argues that to navigate a world characterised by noise and ambiguity, the brain does not simply react to sensory input but actively works to stay ahead of it.
He uses the metaphor of a surfer to explain this process: just as a skilled surfer stays "in the pocket" by remaining slightly ahead of the breaking wave to maintain power and avoid being crashed, the brain "surfs" the waves of sensory stimulation by constantly generating predictions about what will happen next. This predictive mechanism allows the mind to deal rapidly and fluently with an uncertain environment.
Introduction: Guessing Games
In the introduction Andy Clark presents the brain as a prediction-generating system. Rather than passively receiving information from the senses, the brain continually makes guesses about what is happening in the world and compares those predictions with incoming sensory signals.
The central idea is that perception is a kind of “guessing game.” The brain uses prior knowledge and expectations to interpret ambiguous or incomplete sensory input. When its predictions are wrong, the resulting prediction errors prompt it to revise its model, or to act on the world in ways that make sensory input fit its expectations.
Clark connects this predictive process to:
Perception: experience is shaped by expectations as well as sensory data.
Action: we move and intervene in the world to reduce uncertainty and fulfill predictions.
Embodiment: the brain’s predictions are closely linked to the body’s abilities and needs.
Cognition: thinking is not separate from perception and action but part of the same prediction-driven system.
The introduction’s argument is that minds do not simply represent a pre-existing world. They actively anticipate, interpret, and interact with it, allowing organisms to cope efficiently with uncertainty. “Surfing” uncertainty means staying ahead of changing sensory information through continuous prediction and adjustment.
I: THE POWER OF PREDICTION
1 Prediction Machines
1.1 Two Ways to Sense the Coffee
When perceiving the coffee left on my desk, one approach is bottom-up: my brain receives visual signals that build up perception from basic features. This aligns with traditional cognitive models. Alternatively, a prediction-based approach suggests that expectations about the coffee color and shape guide perception, making it a proactive process that refines predictions against incoming sensory information.
1.2 Adopting the Animal’s Perspective
The knowledge required to form predictions comes from prior experiences. Understanding perception involves distinguishing basic sensory data from the richer, meaningful perceptions formed when these data meet expectations. Instead of focusing solely on outer observations, it’s crucial to consider the internal processes of an organism that rely on sensory inputs and energetic stimulations to inform action.
1.3 Learning in Bootstrap Heaven
Prediction-driven learning contrasts traditional supervised learning by allowing systems to learn continuously from the environment rather than relying on pre-categorised data. This approach enhances learning by constantly generating its own teaching signals, making it especially effective in understanding the complexities of sensory data.
1.4 Multilevel Learning
Prediction-driven learning addresses the challenge of representing complex structures. It allows neural systems to infer the causes of sensory inputs efficiently. This process relies on a structure that combines generative models and probabilistic inference, improving the capacity to perceive information in the environment.
1.5 Decoding Digits
Machine-learning systems, like those for recognising handwritten digits, exemplify the utility of probabilistic generative models. By repeatedly adjusting predictions based on input data, these systems can classify diverse information effectively without needing explicit training.
1.6 Dealing with Structure
Prediction-driven learning enables systems to effectively handle structured data by modeling complex relationships. This contrasts with previous symbolic approaches that proved inflexible and displays the adaptability of neural networks in processing structured information.
1.7 Predictive Processing
At the heart of predictive processing is the use of top-down connections to generate predictions that are confirmed or revised based on sensory input. This paradigm posits that perception occurs through a cycle of prediction and error correction, making it a continuous and dynamic process.
1.8 Signalling the News
Predictive coding mechanisms enable sensory neurons to optimise the amount of information by encoding deviations from expected sensory inputs, signalling the most notable changes in the environment. This adaptive process enhances the organism's ability to interpret varying contexts.
1.9 Predicting Natural Scenes
Top-down predictive coding involves a bidirectional flow of predictions and errors across the brain, effectively modeling the rich structure of the sensory inputs. This framework emphasises the role of prediction in mediating perception, allowing for rapid adjustments based on new information.
1.10 Binocular Rivalry
The phenomenon of binocular rivalry illustrates how conflicting sensory inputs lead to alternating perceptions. Predictive processing frames this within a competition between various hypotheses about a scene, highlighting the brain’s active role in making sense of ambiguous stimuli.
1.11 Suppression and Selective Enhancement
Perception relies on minimising sensory prediction error, which involves both suppressing well-predicted aspects and enhancing important features. This duality facilitates efficient processing, enabling the system to prioritise salient information.
1.12 Encoding, Inference, and the Bayesian Brain
Neural representations in a predictive processing framework encode probability distributions, adhering to Bayesian principles where predictions are continuously adjusted based on incoming sensory data. This design equips the brain to handle uncertainty effectively.
1.13 Getting the Gist
Prediction-driven systems rapidly identify gist elements of scenes before fine-tuning details, following a model that acknowledges the importance of immediate context in shaping perceptions.
1.14 Predictive Processing in the Brain
Empirical studies supporting predictive processing reveal how brain responses align with the principles of prediction-driven processing, demonstrating a pattern of prediction errors informing neural activities across varying contexts.
1.15 Is Silence Golden?
Concerns about representing sensory input in light of predictive processing are addressed by acknowledging the interaction between representation and error signaling in neural mechanisms, enriching the understanding of perceptual complexity.
1.16 Expecting Faces
Evidence regarding face processing in brain regions shows that neural responses reflect both predictions and prediction errors, highlighting the complexity of perception beyond simple feature detection.
1.17 When Prediction Misleads
The Hollow Face, is an optical illusion in which the perception of a concave mask of a face appears as a normal convex face because the brain assumes that any face should be convex. This illustrates how our predictive brain assumptions can distort perception, showing that our brain often prioritises prior knowledge over ambiguous sensory information.
1.18 Mind Turned Upside Down
Predictive processing establishes that the brain is an active, anticipatory agent, constantly predicting sensory inputs rather than a passive receiver. This shifts the understanding of perception from a bottom-up approach to one that emphasises prediction and active engagement with the environment. Understanding quantum theory has profound implications for our perception of reality. Just as predictive processing reshapes our grasp of sensory inputs, quantum theory challenges traditional views, suggesting that particles exist in a state of probability until observed. This intersection prompts fascinating questions about the nature of knowledge and the mechanism of perception in both physics and neuroscience.
2 Adjusting the Volume (Noise, Signal, Attention)
2.1 Signal Spotting
Human perception can adeptly isolate specific signals from noise. For instance, when searching for car keys among clutter, our perception adjusts to focus on the target, akin to recognizing shapes like faces or animals in ambiguous patterns. This ability reflects the brain's underlying probabilistic prediction mechanisms, which are shaped by our experiences and expectations.
2.2 Hearing Bing
An experiment shows how expectations can lead to auditory hallucinations. Participants, anticipating a faint sound in a noise track, often "detected" it due to their expectations. This demonstrates the brain's capacity to amplify certain signals while disregarding noise. Similar effects can be seen in sine-wave speech, where exposure to meaningful sentences alters perception. The brain uses top-down (predictive) processing to interpret sensory input, influenced by prior knowledge.
2.3 The Delicate Dance between Top-Down and Bottom-Up
In varying contexts, the brain adjusts its strategy between utilizing prior knowledge and relying on incoming sensory data. Predictive processing (PP) suggests that our brains continuously estimate sensory uncertainty, balancing top-down (expectation-driven) and bottom-up (sensory-driven) influences based on precision-weighting, thus optimising perceptual responses.
2.4 Attention, Biased Competition, and Signal Enhancement
Attention functions as a mechanism to enhance the reliability of sensory information by adjusting the weighting of prediction errors according to their estimated precision. This process not only promotes signals deemed important but also suppresses less pertinent information, thus facilitating better perceptual performance.
2.5 Sensory Integration and Coupling
When multiple sensory inputs are present, the brain integrates them according to their estimated precision, allowing for coherent perception of environmental stimuli. Such coupling between different sensory modalities aids in constructing a unified experience.
2.6 A Taste of Action
Action plays a crucial role in precision-weighted sensory sampling, reinforcing the connection between perception and action. When interpreting sensory input, actions are guided by expectations to reduce uncertainty and confirm predictions.
2.7 Gaze Allocation: Doing What Comes Naturally
In natural tasks, attention distribution is driven by prior knowledge rather than only salience. Learned models dictate gaze allocation to optimise information gathering. Evidence shows that experienced individuals anticipate where relevant stimuli will likely appear.
2.8 Circular Causation in the Perception-Attention-Action Loop
Perception, attention, and action are interconnected processes that influence one another, forming a loop that drives enhanced selective sampling based on systemic predictions and belief validation.
2.9 Mutual Assured Misunderstanding
Misguided expectations can lead to age-old cycles where individuals reinforce incorrect perceptions. For instance, if someone believes another is angry, this bias can shape how they perceive that person's actions and expressions, consequently escalating misunderstanding.
2.10 Some Worries about Precision
Concerns arise regarding whether the predictive processing model adequately addresses certain forms of attention, such as feature-based attention, which captures our senses at unexpected locations due to task relevance rather than spatial cueing.
2.11 The Unexpected Elephant
The difference between neural predictions and subjective agent experiences illustrates how surprising occurrences can arise from high-weighted sensory signals, leading to unexpected perceptions that still align with systemic predictions.
2.12 Some Pathologies of Precision
Disruptions in the mechanisms of precision-weighting may lead to perceptual and belief anomalies, such as hallucinations and delusions in conditions like schizophrenia. These symptoms stem from a breakdown in the balance between top-down and bottom-up processing.
2.13 Beyond the Spotlight
Attention, rather than a simple spotlight mechanism, represents a complex dimension within predictive models of sensory processing. Attention is tightly intertwined with precision expectations, guiding actions to optimise the acquisition of accurate sensory information, thereby enhancing overall perceptual and cognitive experiences.
3 The Imaginarium
3.1 Construction Industries
Perception is a constructive process rooted in predictive mechanisms, which allows beings capable of this perception to imagine and explore their environments beyond mere sensory input. While simpler organisms may react to sensory signals without such elaborate processing, humans and similar beings harness stored knowledge to reconstruct sensory experiences. This generative model can simulate sensory signals and facilitates mental time-travel — allowing for recollection and prediction of events — which ultimately enhances decision-making and choice.
3.2 Simple Seeing
Perception is influenced by prior beliefs that shape our experience of stimuli, as illustrated by the Cornsweet Illusion: the visual system is sensitive to the relative brightness of an area, rather than its absolute brightness. The brain compares the relative brightness of an area to the surrounding areas, this causes it to perceive the brighter area as less bright and the darker area as brighter, than they actually are:
Both sides are the same colour
The brain interprets sensory input through a lens of expectations based on prior experiences, sometimes leading to incorrect perceptions. This predictive mechanism is deemed 'Bayes optimal,' representing an effective way of interpreting the world based on sensory evidence.
3.3 Cross-Modal and Multimodal Effects
Predictive processing explains how context from other sensory modalities influences perceptual responses, even in 'early' sensory areas. Studies reveal that higher-level predictions can shape responses in less stimulated areas, demonstrating a complex interaction of modalities that allows for richer sensory experiences and feedback.
3.4 Meta-Modal Effects
The Visual Word Form Area (VWFA) exemplifies a brain region that tracks abstract representations irrespective of the sensory modality. This area operates as a 'metamodal operator,' enabling predictions about sensory states, including tactile stimuli, demonstrating the brain's ability to integrate diverse inputs.
3.5 Perceiving Omissions
The predictive processing framework effectively accounts for responses to unexpected events, revealing how the brain reacts to omissions in a stimulus. It illustrates how a strong prediction error can arise even in the absence of sensory information, as the brain attempts to reconcile expectations with reality.
3.6 Expectations and Conscious Perception
Expectations fundamentally alter conscious perception, making familiar stimuli more perceptually prominent. The interplay between memory, attention, and prediction underscores how perception can be generally enhanced by expectations, leading to faster conscious awareness of stimuli.
3.7 The Perceiver as Imaginer
Generative internal models enable complex perceptual capabilities and imaginative thought. Creatures capable of rich perception can also generate sensory-like states internally. This duality suggests that perception and imagination are closely intertwined, sharing underlying neural mechanisms for both processes.
3.8 ‘Brain Reading’ During Imagery and Perception
Research supports the notion that perception and mental imagery activate similar brain areas, revealing a significant overlap in neural representations. Experiments show that the brain’s patterns of activity during imagery can mirror those occurring during direct perception, reinforcing the predictive processing model.
3.9 Inside the Dream Factory
Dreaming shares mechanics with perception and imagery but lacks sensory input, leading to less stability and detail. Alterations in neurochemical states during sleep affect predictive error processing and can influence the vividness and coherence of dream experiences.
3.10 PIMMS and the Past
Episodic memory incorporates elements of predictive processing, allowing for the reconstruction of past experiences. The Predictive Interactive Multiple-Memory System (PIMMS) model explains how different memory types interact to optimise the prediction of items and contexts, contributing to a nuanced understanding of recollection and familiarity.
3.11 Towards Mental Time Travel
Mental time travel encompasses recalling past events and anticipating future ones, involving neural substrates that support both functions. This suggests a rely on flexible memory systems that integrate past experiences to inform future predictions, thus enhancing cognitive adaptability.
3.12 A Cognitive Package Deal
The predictive processing model presents a unified framework linking perception, memory, and imagination. This cognitive package highlights how varied mental functions emerge from shared prediction mechanisms, shaping how organisms interact with their environments across various temporal and spatial scales. The exploration of the neural underpinnings of these processes continues to connect understanding with broader embodied experiences.
4 Prediction-Action Machines
4.1 Staying Ahead of the Break
To effectively navigate the continuous sensory stimulus, our ability to predict the future — particularly the anticipated trajectories of our movements — is crucial. This aspect of our cognitive architecture allows us to adjust our actions according to past experiences and desired outcomes. Our predictive machine integrates knowledge of the dynamic environment (like surfing waves) with the expected sensory feedback from our body movements. This self-fulfilling prophecy in action encourages experts to anticipate and actualise desired sensory flows, enhancing their interaction with the world. The interplay of action and agency leads to a deeper understanding of human cognition and might reveal insights into conditions like schizophrenia and autism.
4.2 Ticklish Tales
The peculiar inability to tickle oneself is explained through a model involving a forward model of motor functions that predicts the sensory impact of our self-initiated movements. When we try to self-tickle, our brain anticipates the sensations and reduces our sensitivity to them, which highlights the distinction between externally generated sensations and self-induced ones. This understanding relates to phenomena like force escalation, where individuals perceive their self-generated forces as weaker than those applied externally. Experimental setups demonstrate how delays and variations can increase the ticklishness of self-induced sensations, revealing complexities in our sensory processing.
4.3 Forward Models (Finessing Time)
Forward models are critical for overcoming sensory delays that obstruct seamless motor actions. These models, which connect motor commands to predicted sensory effects, allow for fluid interaction with the environment. Learning and updating these models through experiences enable us to respond quickly and effectively, minimising unnecessary sensory feedback from our own movements while enhancing the perception of external stimuli.
4.4 Optimal Feedback Control
Effective motor control leverages both forward models and inverse models to translate intentions into actions. While forward models predict sensory consequences, inverse models determine the motor commands needed for desired outcomes. Recent developments like optimal feedback control illustrate how ongoing adjustments to actions can occur in real-time, utilising feedback to clarify movements while minimising disturbances.
4.5 Active Inference
The active inference paradigm aligns closely with predictive processing, suggesting that both the perception and motor systems work to minimise prediction errors through a hierarchical structure. Unlike traditional views that separate these functions, active inference indicates they are interconnected, with motor actions arising as predictions of proprioceptive states that guide behaviour towards fulfilling sensory expectations.
4.6 Simplified Control
Active inference proposes a reallocation of roles within cognitive architectures. Action emerges as a natural result of expectations, optimising regarding sensory consequences and bodily dynamics within a generative model. This approach emphasises the importance of embodied experiences in shaping cognitive expectations and directing behaviour.
4.7 Beyond Efference Copy
Current proposals suggest that the concept of outside copy becomes subordinate to the understanding of forward models and corollary discharge. The predictions of sensory outcomes from actions serve directly as motor commands, thereby eliminating the need for separate outside signals, further refining the understanding of how actions are controlled.
4.8 Doing Without Cost Functions
Active inference integrates cost functions into generative models, meaning behavioural outcomes arise from complex expectations rather than predefined values. Consequently, motivations are internalised within expectations, where rewards and pleasures result from interactions rather than serving as direct causes of behaviour.
4.9 Action-Oriented Predictions
The emerging representation systems within a generative model are both predictive and action-oriented, enabling the coordination of perception and action towards minimising sensory prediction error. This convergence emphasises the seamless integration of cognitive processes rather than isolated functionalities, illustrating the co-construction of behaviour driven by action uses.
4.10 Predictive Robotics
In cognitive developmental robotics, the control structures for robot agents are modelled to emerge from interactions with their environments, paralleling human learning processes. Robots engage in motor babbling to establish connections between actions and sensory feedback, resembling human developmental stages, including planning and imitation.
4.11 Perception-Cognition-Action Engines
The predictive processing framework suggests an repetitive relationship between perception, cognition, and action, emphasising that actions stem from simulations of anticipated effects. As prediction models advance, this understanding supports the development of a unified cognitive system that brings together sensory processing and motor control, altering how we interpret cognition in relation to behaviour. Ultimately, the relationship between action and perception becomes an intricate dance aimed at reducing prediction error as organisms adapt to their environments. Artificial intelligence plays a crucial role in enhancing our understanding of predictive models in cognitive processes. By simulating human learning and decision-making, AI can help refine our approaches to sensory processing and motor control. This integration not only advances our knowledge of cognition but also opens new avenues for developing intelligent systems that interact seamlessly with their environments.
5 Precision Engineering: Sculpting the Flow
5.1 Double Agents
The brain operates as a continuous probabilistic prediction machine, actively matching sensory input with expectations and guiding actions based on these predictions. Actions are informed by generative models that encode sensory expectations and understanding others' actions may arise from similar predictive structures. Mirror neurons exemplify this concept, bridging action performance and observation. They fire both when an animal acts and when the animal observes the same action performed by another.
5.2 Towards Maximal Context-Sensitivity
Context sensitivity in perception is illustrated through examples of how predictive expectations bias interpretations of ambiguous stimuli. The predictive processing paradigm enhances this sensitivity by combining hierarchical structures with flexible precision-weighting, making it a fundamental aspect of neural response.
5.3 Hierarchy Reconsidered
Hierarchical organisation in the brain involves reciprocal connections allowing for feedback and feedforward interactions. Findings show that these connections form a complex network that dynamically supports task-specific processing, contradicting the idea of a rigid serial flow of information.
5.4 Sculpting Effective Connectivity
Effective connectivity describes the influence between neural systems shaped by context-specific expectations. Differentiation between structural, functional, and effective connectivity reveals rapid alterations in connectivity based on cognitive tasks, guided by prediction error signals.
5.5 Transient Assemblies
The architecture of predictive processing supports flexible structures called TALoNS (Transiently Assembled Local Neural Subsystems), which can dynamically form based on changing tasks. This flexibility allows for efficient responses in varying contexts.
5.6 Understanding Action
The ability to understand one’s actions and those of others relies on complex predictive mechanisms beyond simple mirroring, with contextual information challenging the interpretation of intentions behind observed actions.
5.7 Making Mirrors
Mirror neurons may not be the sole mechanism for understanding others' actions. Rather, they are products of associative learning, where experience facilitates the recognition of similar actions between agents.
5.8 Whodunit?
Precise distinctions must be made between the predictions governing personal actions versus observational contexts. Contextual cues help adjust proprioceptive predictions to facilitate understanding of others' intentions.
5.9 Robot Futures
Predictive processing allows the potential for mental simulations, granting agents the ability to plan future actions effectively. Robotic simulations demonstrate this capacity, merging generative models with action prediction.
5.10 The Restless, Rapidly Responsive, Brain
The brain operates continually in a proactive manner, utilising both external and internal cues for contextual understanding. Ultra-rapid gist processing aids in efficient recognition of scenes, driven by various neural pathways.
5.11 Celebrating Transience
Predictive processing creates a dynamic cognitive architecture where neural influence is modulated by changing contexts and precision-weighting. This setup enables rapid adaptation to sensory inputs and task demands, establishing the brain as a highly adaptable and fluid system.
6 Beyond Fantasy
6.1 Expecting the World
The brain functions as a probabilistic prediction machine, prompting questions about the relationship between mind and world. This leads to the idea that our perceptions might be a form of 'controlled hallucination' or virtual reality, influenced by complex inner probabilities. However, perceiving accurately is more about guiding action than merely creating internal representations. The essence of understanding our world lies in embodied action and demands of adaptive responses rather than static expectations.
6.2 Controlled Hallucinations and Virtual Realities
Chris Frith suggests that perceptions are models created by the brain that correspond to the world, positioning these perceptions as 'controlled hallucinations.' This concept is echoed by Jakob Hohwy, who emphasizes that perception is indirect, based on the brain's hypotheses about the world, which can sometimes lead to errors, including delusions. Pure perception involves integrating sensory signals and expectations through prediction errors, ultimately reflecting the world in a structured manner that justifies understanding as primarily action-oriented.
6.3 The Surprising Scope of Structured Probabilistic Learning
Prediction-based learning uncovers interacting original causes, enabling effective prediction of sensory signals. Hierarchical Bayesian Models (HBMs) illustrate how abstract structure can emerge from raw data without extensive innate knowledge, thus reconfiguring discussions of nativism versus empiricism. Through multistage learning, systems can form high-level schemas that simplify the learning process, influencing perceptions based on broad expectations extracted from environmental interactions.
6.4 Ready for Action
Traditional models of cognition often present a linear sense-think-act paradigm, which fails to capture the complexities of real-time decision-making. The interactive competition hypothesis proposes that multiple potential actions are computed in parallel, leading to proactive responses based on context and sensory information. This narrative challenges classical separations in cognitive processing, suggesting a more integrated view of perception, cognition, and action.
6.5 Implementing Interactive Competition
Utilizing the predictive processing framework, affordance competition emerges as a consequence of probabilistic prediction, which continuously informs the selection and execution of action. The distinctions usually made between perception, cognition, and action blur as the brain actively prepares multiple potential actions based on the current context, navigating the complexities of real-world interactions.
6.6 Interaction-Based Joints in Nature
Probabilistic prediction-driven learning allows us to see past superficial sensory noise and reveals the interacting structures of the environment. Our perception is influenced by our action repertoire, which enables efficient processing of relevant features within our environment. The integration of action and perception fosters an understanding of the world that is deeply tied to our biological needs and adaptive functions.
6.7 Evidentiary Boundaries and the Ambiguous Appeal to Inference
Hohwy argues that prediction error minimisation creates a boundary that confines the mind's inferential processes within the skull, suggesting that the mind may be isolated from the external world. This raises questions about how well the mind represents reality, especially regarding the application of evidence in cognitive processing, indicating a fundamental tension between traditional embodied cognition and the inferential stance.
6.8 Don’t Fear the Demon
Hohwy links the concept of inferential seclusion to global scepticism, asserting that the brain might simulate an embodied experience while being misled by sensory signals. However, this scepticism does not negate the insights of embodied cognition. Rather, it emphasises the crucial interplay between action and perception in real-time scenarios.
6.9 Hello World
The predictive processing framework can facilitate a structured perception of language, where meaning is derived from the sound stream amidst noise. Although perceptions require prior knowledge for clarity, they are not mere fantasies but real representations of interaction with the environment.
6.10 Hallucination as Uncontrolled Perception
Hallucinations can be seen as uncontrolled perceptions where sensory prediction mechanisms fail. Such disturbances may arise from malfunctioning neural processes, leading to erroneous beliefs and perceptions that do not align with the external world.
6.11 Optimal Illusions
Illusions may reflect optimal perceptual strategies rather than failures, allowing behavioural success despite occasional inaccuracies. Illusory experiences contribute to a coherent computational strategy tailored to navigate an uncertain sensory environment.
6.12 Safer Penetration
Perceptual systems are subject to top-down influences, and while they may succumb to biases from cognitive processes, they are still adequate in their calibration to establish reliable interaction with the external environment.
6.13 Who Estimates the Estimators?
Severe mental disruptions disrupt the balance between sensory input and predications, complicating the establishment of reliable perceptions. The dynamic relationship between these factors can lead to cycles of false beliefs and experiences, particularly in forms of psychosis.
6.14 Gripping Tales
Ultimately, perception as a probabilistic prediction process is an active system that adapts to accommodate changing sensory states. This perspective highlights that our interactions with the world are shaped by our needs and the affordances our environment provides, leading to a rich, meaningful engagement with our surroundings rather than a mere virtual reality.
7 Expecting Ourselves (Creeping Up On Consciousness)
7.1 The Space of Human Experience
The chapter begins by reflecting on how human experience can be understood through the lens of predictive processing. The exploration covers the mechanisms of perception, imagination, action, and reasoning, emphasising the dynamic role of uncertainty in neural processing. The goal is to connect theoretical frameworks with the complexities of human experience, revealing the intricacies of the mind's functioning in various states.
7.2 Warning Lights
This section introduces scenarios where human experience is disrupted and how these disruptions can be explained through predictive processing principles, particularly concerning prediction errors and their reliability. Problems may arise from overestimating the precision of evidence, leading to distorted perceptions and delusions.
7.3 The Spiral of Inference and Experience
Here, the focus is on how delusions and hallucinations, especially in schizophrenia, stem from abnormal prediction errors. It highlights the self-reinforcing nature of these erroneous beliefs, fostering bizarre interpretations of sensory experiences.
7.4 Schizophrenia and Smooth Pursuit Eye Movements
This section discusses eye movement anomalies in schizophrenia, noting that difficulties in tracking moving objects are linked to disturbances in predictive processing. These findings relate back to broader themes of cognitive disturbances associated with the condition.
7.5 Simulating Smooth Pursuit
Adams et al.’s work on simulating smooth pursuit eye movements is outlined, suggesting that predictive processing models can illuminate the rationale behind differences in response patterns between neurotypical and schizophrenic subjects.
7.6 Disturbing the Network (Smooth Pursuit)
The focus shifts to the implications of altered precision in prediction errors on eye movement and associated cognitive behaviours. These disturbances are shown to affect learning and performance in individuals with schizophrenia.
7.7 Tickling Redux
This section revisits findings about self-tickling in schizophrenia, exploring the interconnectedness of sensory experiences, motor control, and the emergence of delusions of agency.
7.8 Less Sense, More Action?
Discusses how impairments in sensory processing precision affect movement generation, emphasising the role of interoceptive and proprioceptive signals in the execution of actions.
7.9 Disturbing the Network (Sensory Attenuation)
The consequences of disrupted sensory attenuation on movement and perception are examined, demonstrating the relationship between sensory input precision and action initiation.
7.10 ‘Psychogenic Disorders’ and Placebo Effects
This section reviews functional motor and sensory symptoms attributed to psychogenic causes, proposing theories on how expectations can shape physical experiences and reflect predictive processing mechanics.
7.11 Disturbing the Network (‘Psychogenic’ Effects)
Explores how disturbances in precision-weighting mechanisms can lead to unexplained sensory and motor symptoms and connect to the development of erroneous beliefs about bodily agency.
7.12 Autism, Noise, and Signal
The relationship between predictive processing and the non-social symptoms of autism is discussed, with a focus on how variations in sensory processing precision impact behaviour and experience.
7.13 Conscious Presence
Seth et al.'s theory on the feeling of conscious presence is presented, arguing that this sensation arises from effective integration of interoceptive signals and prediction accuracy.
7.14 Emotion
The role of interoception in emotional experiences is examined. Emotion is described as the result of interplay between physiological states and top-down predictions that are shaped by context.
7.15 Fear in the Night
Pezzulo elaborates on irrational fears, like fear at night, suggesting that the interplay of interoceptive and exteroceptive prediction influences how we interpret sensory data in certain contexts.
7.16 A Nip of the Hard Stuff
The chapter concludes by synthesising insights from predictive processing and emphasising the potential for a unified understanding of human experience. This reflects a promising future for the field of computational psychiatry and its potential applications in understanding consciousness and mental health.
8 The Lazy Predictive Brain
8.1 Surface Tensions
The chapter begins with a reference to the documentary "Fast, Cheap, and Out of Control," highlighting the evolution of behaviour-based robotics, notably by Rodney Brooks. Brooks challenged traditional "symbolic, model-heavy" approaches, promoting simpler, more efficient responses suitable for complex environments. Although some aspects of Brooks’ robots were limited in scope, they contributed significantly to understanding how intelligent agents can utilise their bodies and environments effectively. Predictive Processing (PP) emerges as a framework that combines simple fast responses and complex strategies, advocating for efficient problem-solving through an integration of mind, body and environment.
8.2 Productive Laziness
"Productive laziness," a concept borrowed from Aaron Sloman, suggests that organisms often employ practical heuristics rather than optimal solutions, demonstrating the effectiveness of being “good enough” under constraints of time and cognitive resources. This notion of "satisficers" led to the concept of bounded rationality, which has implications for decision-making and judgment.
8.3 Ecological Balance and Baseball
The chapter emphasises the ecological efficiency in actions, likening it to bipedal locomotion in robots versus biological beings. Nature's designs exploit passive dynamics, resulting in efficient movement that contrasts sharply with energy-intensive robotic walking. The ecological balance principle showcases how sensory and motor systems should match the complexity of the environment, supporting efficient problem-solving across agents.
8.4 Embodied Flow
Embodied cognition dismisses the rigid sequence of perceiving, thinking, and acting. Instead, real-world interaction blends these processes, as evidenced by a study of how individuals copy patterns using rapid eye movements. This intertwining of perception and action suggests that the brain operates not as a centralised reasoning engine but as a system adept at seamlessly engaging with the environment for effective action.
8.5 Frugal Action-Oriented Prediction Machines
Predictive processing isn't just about understanding the world but entails guiding action. Agents use predictive mechanisms to engage in affordance competition, developing representations that prioritise effective control over sensory experience, making neural processes fundamentally action-oriented.
8.6 Mix ‘n’ Match Strategy Selection
The relationship between predictive processing and decision-making frameworks elucidates how agencies toggle between model-based and model-free strategies based on context. Rather than fixed systems, these strategies are intermingled, adjusting to the task demands and uncertainties involved in decision-making processes.
8.7 Balancing Accuracy and Complexity
Bayesian principles advocate for agents to maximise prediction accuracy while minimising model complexity. This balance can lead to the emergence of a unified architecture combining fast heuristic strategies with more complex deliberation, dynamically adjusted by contextual demands and resource attention.
8.8 Back to Baseball
In reference to the outfielder's problem in baseball, the predictive framework illustrates how agents utilise environmental cues to minimise prediction error and effectively engage with dynamic situations. Adaptive learning contributes to a flexible response strategy, adjusting based on context.
8.9 Extended Predictive Minds
The chapter extends the notion of cognition beyond the brain to include environmental and social factors, asserting that real-time cognitive processes can leverage external resources for improved predictions and actions. Effective problem-solving emerges from interactions among cognitive systems, the body and the world.
8.10 Escape from the Darkened Room
Friston's Darkened Room Puzzle raises concerns about predictively oriented behaviors leading to inactivity. The chapter counters this notion by demonstrating that evolved organisms are inherently motivated to seek novelty and exploration, driven by complex adaptive expectations formed through evolutionary processes.
8.11 Play, Novelty, and Self-Organised Instability
Novelty-seeking behaviour challenges the predictivity model, suggesting that exploration can be a structural part of cognitive functioning. Various studies indicate that humans and infants actively seek moderate complexity, contributing to self-organised behaviour and learning throughout their lives.
8.12 Fast, Cheap, and Flexible Too
In dynamic environments, creatures deploy a range of strategies to address varying situational demands. These strategies, from simple heuristics to complex reasoning, reflect the adaptability of the predictive brain in managing both cognitive and physical resources efficiently. The chapter concludes by affirming that the predictive brain is aimed at sustainable engagement with the world, optimising outcomes by minimising unnecessary cognitive load.
9 Being Human
9.1 Putting Prediction in Its Place
The chapter begins with the assertion that our neural systems are designed to support embodied actions through complex interactions of perception and action. This creates a dynamic self-organisation process that minimises prediction errors, ultimately enabling humans to adapt and thrive within their sociocultural environments. The interplay between neural dynamics and social structures is pivotal for understanding human cognition.
9.2 Reprise: Self-Organising around Prediction Error
The role of prediction error in neural self-organization is emphasised. Systems learn to predict sensory inputs through a probabilistic generative model, which reveals complex interdependencies in different contexts. This process supports adaptive learning, allowing for a variety of predictive routines adaptable to new situations.
9.3 Efficiency and ‘The Lord’s Prior’
Efficiency in cognitive systems is contrasted with redundancy, highlighting that simpler models (like the Optical Acceleration Cancellation) can effectively minimise prediction errors. Efficiency is viewed as crucial for maintaining a balance between sensory input and appropriate behavioural response without overfitting data.
9.4 Chaos and Spontaneous Cortical Activity
The concept of synaptic pruning is discussed as a method for refining neural models by removing irrelevant connections. The chapter also explores spontaneous cortical activity as a form of exploration that aids in creative problem-solving by enabling mental simulations and cognitive flexibility.
9.5 Designer Environments and Cultural Practices
Human cognition is not unique but is enhanced by cultural practices and language. These elements foster a structured environment where prediction-driven learning thrives, leading to the development of complex concepts and ideas that transcend basic biological adaptations.
9.6 White Lines
The importance of simple environmental modifications that prompt cognitive shortcuts is discussed. These changes allow for easier behavioural navigation, enhancing the utility of prediction error minimisation through culturally altered environments.
9.7 Innovating for Innovation
Cultural learning processes are explored as innovations arising from social interaction rather than just biological evolution. Practices like reading exemplify how culture can reshape cognitive processes, creating tools that facilitate learning and knowledge transfer.
9.8 Words as Tools for Manipulating Precision
Language serves dual functions, aiding communication while actively shaping our thoughts. Evidence suggests that words can enhance perception and manipulate the precision of predictive processes, thereby improving cognitive outcomes.
9.9 Predicting with Others
The mutual prediction facilitated by social interactions highlights how language and shared understanding enable seamless communication and collaborative learning. This fosters a reciprocal system where individual actions are refined through social cues.
9.10 Enacting Our Worlds
The chapter concludes by discussing how humans actively construct the worlds they inhabit. This process of 'enactment' emphasises the reciprocal relationship between organisms and their environments, shaped by actions that guide perceptions and vice versa.
9.11 Representations: Breaking Good?
The tension between predictive processing and enactivism regarding internal representations is addressed. While PP employs representations to manage action and perception, these models are dynamic and context-sensitive, focusing on engagement with the environment rather than mere depiction.
9.12 Prediction in the Wild
Finally, the chapter argues that our neural systems are fundamentally geared for action, emphasising the importance of prediction in shaping our experiences and interactions. The interplay of cultural practices, technology, and neural processes shapes our cognitive landscape, marking a significant challenge for cognitive science in the twenty-first century. Understanding brain activity is crucial for deciphering the complexities of human cognition. This interplay between brain activity, cultural practices, and social interactions not only shapes our predictive models but also influences how we process information. As we explore the depths of neural dynamics, we uncover the significant role that environmental factors play in enhancing cognitive functions.
10 Conclusions: The Future of Prediction
10.1 Embodied Prediction Machines
Predictive processing (PP) presents a view of the brain as a dynamic system closely linked to embodied cognition, where action and perception are interwoven. This ongoing interaction reflects how perception, understanding, reasoning, and imagination work together, characterised by a flow of neural activity that adapts to uncertainty. The brain acts not only as a generator of inferences but as an active participant in engaging with the world, constantly refining its predictions to maximise adaptive behaviour. By utilising multilevel probabilistic generative models, this predictive framework enables organisms to discern meaningful signals from noise and navigate sensory uncertainties efficiently. Both perception and action arise from this predictive machinery, working in tandem to shape interactions with the environment and informing cognitive strategies that optimise actions based on bodily and contextual factors.
10.2 Problems, Puzzles, and Pitfalls
While this vision of the embodied mind represents significant progress, several challenges remain. First, there is a need to explore various approximations and representational forms that the brain employs for complex problem-solving. Understanding the specific strategies the brain uses for probabilistic inference is essential for correlating theoretical models with actual cognitive processes. Second, further investigation into diverse architectural approaches is needed to fully appreciate alternative designs for how predictions and sensory information might interconnect within the brain. Third, there is a call for extending these insights into higher cognitive functions including long-term planning and social cognition, which remain less clearly outlined in terms of generative models. Lastly, a conceptual concern arises regarding the potential risk of reintroducing a Cartesian perspective, suggesting that a heavy reliance on internal models may isolate cognition from the body and environment. The framework advocated here emphasises a reciprocal relationship between inner neural processes and outer environmental stimuli, suggesting a more integrated understanding of cognition that aligns action, perception, and prediction. Central to the discussion of cognitive processes is the concept of probabilistic reasoning. This approach allows individuals to make informed decisions by estimating the likelihood of various outcomes based on prior experiences and available information. Understanding how the brain employs probabilistic reasoning can enhance our insights into learning, decision-making, and problem-solving in both everyday life and complex scenarios.
Themes
Challenge to the traditional notion of the "I"
The traditional notion of the “I” assumes that each person possesses a stable, unified, and independent self. According to this view, beneath our changing thoughts, emotions, and experiences lies a consistent inner identity that remains essentially the same over time. However, many philosophical traditions have challenge this assumption by suggesting that the self is not permanent or independent.
Buddhist philosophy, for example, argues for the idea of anattā, or “no-self.” It claims that what we call the self is actually a collection of constantly changing physical sensations, emotions, perceptions, memories, and thoughts. David Hume made a similar argument in Western philosophy, claiming that when he looked inward, he could not find a permanent self—only a series of passing experiences.
Other thinkers challenge the “I” by emphasising the influence of unconscious forces, society, language, and culture. Nietzsche viewed the self as a collection of competing drives rather than a single unified subject. Psychoanalysis also suggests that people are not always aware of the forces shaping their thoughts and actions. Meanwhile, poststructuralist philosophers argue that identity is formed through language, social expectations, and relationships with others.
Surfing Uncertainty explores the intersection of consciousness, biology, and the nature of the self. A central theme is the challenge to the traditional notion of the "I", with Clark arguing that the mind is not a centralised entity locked inside the skull, but rather a distributed system. He proposes that our cognitive processes extend into our environment and the tools we use, suggesting that the boundaries between the internal mind and the external world are fluid and arbitrary.
The book also heavily emphasises predictive processing, framing the brain not as a passive receiver of sensory information, but as an active "prediction engine." Through this lens, Clark examines how the mind constantly generates hypotheses about the world and updates them based on sensory feedback. This theme highlights the tension between our internal models of reality and the actual state of the environment, suggesting that perception is essentially a form of controlled hallucination.
Additionally, Clark delves into the relationship between embodiment and cognition. He argues that the physical body and its movements are not merely peripherals for the brain, but are integral to how we think and perceive. By examining how we interact with complex environments — such as a surfer navigating a wave — he illustrates that intelligence emerges from the dynamic coupling of the brain, the body, and the world.
Top-down processing
In Surfing Uncertainty, Andy Clark describes top-down processing as the way the brain’s expectations, prior knowledge, and predictions influence perception. Rather than simply receiving sensory information and building an understanding of the world from the ground up, the brain continuously anticipates what it is likely to encounter. These predictions shape how incoming information is interpreted.
Clark contrasts this view with a purely bottom-up account of perception. In a bottom-up model, sensory data enters through the eyes, ears, or other senses and is gradually combined into meaningful objects and events. For Clark, however, sensory input is always interpreted in relation to higher-level expectations. The brain is constantly asking, in effect, what is most likely to have caused the signals it is receiving.
This process can be understood through prediction error. The brain makes predictions at different levels, from broad expectations about a situation to detailed expectations about colors, sounds, shapes, or movements. Incoming sensory information is then compared with those predictions. When the information does not match, the difference produces a prediction error, which can lead the brain to revise its expectations or to focus more closely on the sensory evidence.
Top-down processing therefore does not mean that perception is based only on imagination or personal belief. Instead, perception results from an interaction between prediction and sensory evidence. For example, context can help someone interpret an unclear sound or image. A person who knows the rules of a sport may perceive the movements of players more meaningfully because their prior knowledge allows them to anticipate what is likely to happen.
For Clark, top-down processing also connects perception with action. Organisms do not merely wait for information from the world; they move, explore, and act in ways that help confirm their predictions or reduce uncertainty. Perception is therefore an active process in which the brain uses prior knowledge to interpret the world while continually correcting itself through sensory feedback.
Embodied cognition
For Andy Clark, embodied cognition is the idea that thinking is not produced by the brain alone. Cognition depends on the continuous interaction between the brain, the body, and the environment. The body’s shape, movements, sensory abilities, and practical relationship with the world all contribute to how an organism perceives, thinks, and acts.
In Surfing Uncertainty, Clark connects embodied cognition with predictive processing. The brain continually predicts what will happen, but the body helps test and fulfill those predictions through movement and perception. For example, reaching for an object is not simply an action that follows thought; the movement itself helps gather information about the object and makes the task easier to complete.
Clark also emphasises that action can reduce the brain’s uncertainty. Instead of passively waiting for clearer sensory information, an organism can move closer, change its angle of view, touch an object, or manipulate its surroundings. These bodily actions alter the incoming sensory signals and help the brain determine what is happening.
The environment is therefore an active part of cognition. Objects, tools, symbols, and social surroundings can support mental activity by storing information or making tasks easier. A written note, for example, can function as an external memory, while a calculator can perform operations that would otherwise have to be carried out internally.
Clark’s view argues that cognition emerges from the cooperation of brain-based predictions, bodily abilities, and environmental resources. Perception, thought, and action form a continuous cycle in which the organism uses its body and surroundings to navigate uncertainty.
Prediction error
Prediction error is the difference between what the brain predicts and what actually happens. When a prediction error occurs, the brain adjusts its internal model so that it can make better predictions in the future. In this way, learning is not simply the accumulation of facts, it is the gradual improvement of the brain’s ability to anticipate patterns in the world.
Clark uses language learning as an example. A person who repeatedly tries to predict the next word in a sentence will gradually learn grammatical patterns, meanings, and regularities. Even without being explicitly taught every rule, the learner can develop knowledge by noticing which predictions succeed and which fail. Prediction therefore helps the mind build up more complex understanding from repeated experience.
Learning takes place at several levels of the brain. Lower levels may learn to predict basic sensory features such as lines, sounds, or movements, while higher levels learn to predict more complex events, objects, and situations. These levels interact continuously: higher levels provide expectations, while lower levels send prediction errors upward when sensory input does not match those expectations.
For Clark, the body and environment also contribute to learning. People learn not only by changing their internal beliefs but also by acting, exploring, and modifying their surroundings. Moving closer to an object, testing a possibility, or using a tool can produce clearer sensory information and reduce uncertainty.
Thus, learning through prediction is a central part of Clark’s account of cognition. The mind becomes more intelligent by building generative models of the world, using them to anticipate experience, and revising them through prediction error. In short, we learn by trying to predict what will happen and using our mistakes to improve future predictions.
Clark’s theory of learning compared to the AI learning process
AI learning systems and Andy Clark’s learning theory are similar because both emphasise prediction, feedback, and the correction of errors. An AI system learns by making a prediction, comparing it with the correct answer or with feedback from its environment, and then changing its internal parameters to improve future predictions. Similarly, Clark’s theory of predictive processing suggests that the brain constantly generates predictions about the world and updates them when sensory information does not match those predictions.
The main difference is that AI learning is usually described as a computational process involving data, algorithms, and parameter changes. For example, a neural network may learn to recognize objects by processing thousands of labelled images and adjusting its weights when it makes mistakes. Its goal may be to minimise error, maximise accuracy, or receive a higher reward. Clark’s theory, however, describes learning as a continuous process in which an organism predicts what will happen, senses the world, and acts to reduce the difference between its expectations and reality.
Clark also places much greater importance on the body and the environment. Many AI systems, especially language models, learn from data without having a body or directly interacting with the world. They can identify patterns in language, but they do not necessarily experience objects, movement, or physical consequences. In Clark’s view, cognition is embodied: the brain, body, and environment work together. Actions such as looking, moving, touching, or using objects help a person learn and understand the world.
Another important difference concerns the location of intelligence. Traditional AI tends to treat intelligence as something contained within the computer system. Clark’s extended-mind theory argues that thinking can also depend on external tools and surroundings. A person using notes, a calculator, a map, or a smartphone may be using those objects as parts of the cognitive process. Some modern AI systems that use external memory, search engines, sensors, or tools resemble this idea because their abilities are distributed across the model and its environment.
AI learning systems apply specific mathematical methods to learn patterns and improve performance, whereas Clark’s theory provides a broader explanation of cognition. Both involve prediction and error correction, but Clark’s approach emphasises active engagement with the world, bodily experience, and the use of external resources. An AI system would be more similar to Clark’s theory if it could perceive its environment, act within it, learn continuously from the consequences of its actions, and use tools as part of its thinking.
No comments:
Post a Comment