An introduction to the predictive brain hypothesis, published in 2023. It explains perception, pain, emotion, mental illness, and bias through the view that the brain makes predictions, corrects errors, and changes sensory input through action, rather than passively receiving sensations.

Preface: Shaping Experience

The author describes experiencing phantom vibration syndrome: feeling a phone vibrate when it has not. It is known to be associated with stress. In a 2012 survey of university students, 89% had experienced it, with medical interns experiencing it particularly often.

A mobile phone is always in our pocket and vibrates frequently, causing stress. Modern people therefore unconsciously predict phone vibrations in stressful situations. Since predictions and expectations strongly influence experience and our perception of reality, we “experience” phantom vibrations. This is the book’s subject: the predictive brain hypothesis.

Traditional view: Perception is passive reception of sensory information.
More recent view: Predictions and expectations influence perception; top-down processing exists.
Predictive-processing view: Predictions and expectations are central to perception. Sensory information corrects prediction errors and improves predictions. Moreover, actions consistent with predictions actively influence subsequent perception.

If the brain strongly predicts a particular sight, sound, or feeling, that prediction “shapes” what we see, hear, and feel. In this sense, every human experience is partly a hallucination. Experience without prediction is as impossible as surfing without water. (This seems to borrow the metaphor of his earlier Surfing Uncertainty. —ak)

Chapter 1. Unboxing the Prediction Machine

The brain is a prediction machine, evolved to make and update endless predictions about the world based on actual sensory information and expectations. Differences between predictions and sensory information act as prediction-error signals that correct predictions, producing hallucinations during the brief correction interval.

Prediction applies not only to external sensations but also to perceptions of internal sensations such as pain, hunger, anxiety, and depression. This means we can sometimes change how we feel by consciously or unconsciously adjusting our predictions.

David Marr’s bottom-up model—extracting simple visual features and combining them recursively to perceive the world—is still useful, but differs from the brain’s actual structure. Its wiring carries far more top-down than bottom-up signaling, a design that long remained puzzling.

The brain runs a simulation of the world, updating it based on discrepancies between predictions and perceptual information. Top-down neurons convey the simulation’s predictions to the sensory cortices. For the brain, only the discrepancy between prediction and sensory information is new, so there is no need to send all information upward.

On a radio with poor reception, familiar lyrics are audible, whereas unfamiliar lyrics are almost impossible to understand. The audio signal has the same quality, yet familiarity changes the perceptual experience—another good fit for the predictive-brain hypothesis.

Moreover, as embodied agents, humans continually and actively acquire needed perceptual information through bodily movements, such as moving an ear toward a sound. (It feels like a combination of machine learning’s online learning and active learning. —ak)

According to P. Wallisch 2017, the 2015 dress controversy can also be explained by this hypothesis. Day-active people are accustomed to natural light and interpret it as gold and white; night-active people are accustomed to artificial lighting and interpret it as black and blue. Daily routines establish expectations about illumination, affecting color perception.

How does the brain learn to predict the world? 1) Its initial wiring is largely the result of evolutionary design. 2) The rest is self-supervised learning: discrepancies between the continual stream of sensory input and the brain’s predictions are used to improve subsequent predictions.

Chapter 2. Psychiatry and Neurology: Closing the Gap

A 15 cm nail pierced a man’s shoe. He reported extreme pain, but removing the shoe at the hospital revealed that it had passed between his toes, leaving him uninjured (see somatic amplification). Vivid visual input caused the brain to predict pain, producing a painful experience. Signals from nociceptors are only one of many causes of pain experience.

Nociceptive pain originates in signals from nociceptors about bodily damage; neuropathic pain originates in damage to sensory nerves. Nociplastic pain occurs without an obvious problem such as tissue damage. Predictive processing offers clues about nociplastic pain.

There seems to be a mechanism determining how much importance to assign to prediction-based top-down signals versus sensation-based bottom-up information when they differ. On this hypothesis, the nail case occurred because greater weight was assigned to visual information than to nociceptive information, favoring top-down signals predicting pain.

Various placebo and nocebo effects work similarly. For example, injecting a placebo produces a stronger effect than administering it as a pill, because patients tend to regard injections as more powerful medical interventions than oral medication.

A feedback loop can also form: predicted pain combines with weak nociceptive signals to amplify perceived pain, producing top-down signals that predict pain more strongly. Once formed, this loop can persist, and may contribute to chronic pain.

Traditional psychiatry groups symptoms and chemical changes in the brain by correlation and uses those groupings for diagnosis and prescriptions. Computational psychiatry seeks to address underlying causes rather than symptoms, hypothesizing that neurodiversity and various psychiatric conditions arise from differences in how the brain weights information.

Motor impairment, paralysis, blindness, and so forth occurring without physiological abnormalities are functional disorders, difficult to explain traditionally. One patient reported blindness despite responding normally, including making eye contact with the doctor (see functional visual loss). It emerged that the patient had experienced light-induced migraines and spent a long time in a dark room. The suspected problem was the brain’s persistent prediction of an absence of light. Vision fully recovered through interventions that persuaded the patient that they were actually seeing despite not consciously realizing it, and activated visual centers through TMS (stimulating neurons with strong magnetic fields). (Many similar successes have been reported. See Seeing again: treatment of functional visual loss.)

The brain seems to have a mechanism for estimating the precision of different information and assigning different weights. The author argues that this is what we call attention. Various aspects of functional disorders can be understood as malfunctions of this weighting mechanism operating as self-fulfilling predictions.

There is a spectrum between heavily weighting sensory information and heavily weighting predictions. It is therefore natural for different people to show different symptoms even with the same physiological condition. Functional disorders may occupy one extreme of this spectrum (heavily weighting predictions).

Autism spectrum disorder is thought to arise from excessive weighting of sensory information relative to top-down predictions. The comparatively weak role of prediction makes it difficult to integrate sensory information meaningfully into “useful hallucinations,” producing symptoms such as difficulty reading subtle changes in others’ expressions.

Because sensations are processed intensely, people try to minimize information from their surroundings, reducing interaction with other people—the most complex part of those surroundings. Repetitive hand movements are an unconscious attempt to generate controllable, regular sensory information. Developing great expertise in a particular field is also an effort to reduce surprises.

Predictive processing explains schizophrenia as arising when prediction-error signals are generated despite no discrepancy between prediction and sensation, and then given excessive weight. Strong signals that predictions are wrong lead to increasingly bizarre revisions of the model (or simulation) of the world.

The predictive-processing hypothesis offers a consistent explanation of chronic pain, functional disorders, various mental illnesses, and neurodiversity in terms of differences in how the brain weights predictions and sensory information. It can transcend the mental–physical and psychiatric–neurological dualisms or dichotomies, suggesting directions for better medical interventions.

Chapter 3. Action as Self-Fulfilling Prediction

Prediction is closely connected to action. The brain predicts an action’s outcome and guides action toward minimizing errors in that prediction. Unlike the traditional model’s separate stages of “perception → formation of mental representations → action,” prediction (mental representation?) continuously guides perception and action. Perception and action are tightly coupled. (See Action as self-fulfilling prediction.)

This idea continues the ideomotor theory of H. Lotz in nineteenth-century Germany and William James in the United States: an idea of the outcome of an action elicits that action.

If a discrepancy between the predicted outcome and current sensations simply updates the prediction to fit the sensations, there is no need to act. For action to occur, predictions must therefore receive greater weight. This model is already widely used in sports coaching (The Inner Game of Tennis! —ak): focus on the desired outcome rather than specific movements.

Robotics also offers cases of improved performance through active sensing rather than the traditional sense–compute–act cycle. Here, sensing is part of active behavior that supplies appropriate information at the right time, rather than a separate passive stage.

All kinds of “training” are closer to learning how to predict than how to act. Muhammad Ali expressed this well: “If my mind can conceive it and my heart can believe it, then I can achieve it.”

Long-term behavior (learning to surf, for example) works similarly. Counterfactual predictions (“if…, then…”) reduce prediction errors, extending predictions further through time. Whatever individuals consciously feel, the predictive hypothesis holds that they make “realistic but optimistic” predictions that can fulfill themselves.

Explaining long-term behavior this way is neat, but how can we assess the claim’s validity? Here are the book’s references:

Friston, K. J., et al, “Deep Temporal Models and Active Inference” (2017)
Van de Cruys, et al., “Controlled Optimism” (2020)
Friston, K., et al., “The Anatomy of Choice” (2014)
Friston, K., et al., “Active Inference, Homeostatic Regulartion and Adaptive Behavioral Control” (2015)

Chapter 4. Predicting the Body

The dark-room puzzle: “If we act to reduce prediction errors, why not enter a dark room and sit still?” Predictions serve to keep the organism alive. This requires predicting internal bodily states too: drinking before dehydration, for example.

Homeostasis and allostasis are important here. Homeostasis maintains a fixed state; allostasis changes the target state to suit the situation and then maintains it. Early cybernetics focused on homeostasis, but allostasis is more encompassing and important.

(We often say “cold-blooded” animals lack a mechanism for maintaining body temperature and therefore seek sun or shade. But from the perspective of distributed functional decomposition, moving the body to a suitable environment is itself a temperature-regulation mechanism. The mechanism is different, rather than absent. —ak)

Homeostasis and allostasis require information from inside the body (heart rate, blood sugar, temperature, etc.) as well as outside it. Interpretation of internal information is also influenced by predictions. Internal states may change to fulfill predictions, and weighting mechanisms determine which information matters and how much (see precision variations).

Survival requires pursuing positive outcomes as well as avoiding negative ones. Lower prediction errors encourage new attempts; higher errors discourage them. Improving predictive skill requires an environment that is “appropriately difficult,” rather than too easy or too hard to predict. Agents actively seek such environments.

(I keep thinking of reinforcement learning’s exploration–exploitation dilemma and active learning. Everyday parallels include seeking fun when bored and reduced creativity under stress. —ak)

Emotions and feelings are deeply related to the various ways prediction operates on external and internal information. Lisa Feldman Barrett’s How Emotions Are Made introduces the metaphor of body budgeting.

Animals budget essential resources such as water, salt, and glucose. Drinking when thirsty brings an immediate feeling of relief, although it actually takes about 20 minutes for water to enter the body and exert the desired effects. Since thirst itself is an anticipatory feeling created by prediction, relief is anticipatory too.

According to Barrett, all emotions result from body budgeting. The idea that internal bodily information influences emotion goes back to William James in 1890. W. J. held that bodily changes associated with fear (sweating, trembling) themselves constitute fear. This view has limitations, however, such as being unable to distinguish fear from anxiety.

On the predictive brain hypothesis, emotions and feelings are perceptions integrating exteroceptive information, interoceptive information, proprioceptive information about bodily position, and predictions (am I exercising, or about to have a heart attack?). The anterior insular cortex appears to perform this role.

Contrary to the traditional division between cortex and subcortex, with the cortex responsible for “higher cognition,” the two are closely intertwined. The basal ganglia, for example, connect with the cortex through five recurrent circuits. These close cortical–subcortical connections allow brain, body, and world to participate equally in generating thought, experience, and action.

L. F. Barrett interprets depression as an allostatic disorder: a problem in predicting bodily energy. It underestimates or overestimates present or future energy needs, repeatedly alternating bursts of enthusiasm and energy with sudden fatigue. This also explains why chronic depression accompanies problems with sleep, metabolism, and immunity.

Chronic depression and anxiety share strong resistance to updating predictions formed through past negative events, even when positive events occur. Small, ambiguous signals—expressions, voices, gestures—come to be interpreted increasingly negatively. This can ultimately create a vicious cycle of behavior resembling “entering a dark room.”

Another example of body–emotion–prediction interaction is the aesthetic chill: intense emotional elevation while appreciating art or making a scientific discovery, usually accompanied by goosebumps and hair standing on end.

According to the salience detection hypothesis, this occurs when information resolves an important uncertainty. Music often elicits it because it is constructed to create expectations and uncertainty along multiple dimensions early on, then resolve them at crucial moments.

Karin Kukkonen, a professor of comparative literature at the University of Oslo, calls these “probability designs.” Novels, drama, and films are deliberately designed to interact with predicting brains in reliable ways.

The Fluid Interface Group at the MIT Media Lab demonstrated a device that induces aesthetic chills by artificially producing the bodily temperature changes accompanying them. Another example of the brain using bodily information to interpret a situation.

Interlude. The Hard Problem: Predicting the Predictors?

How does the predictive brain hypothesis explain qualia? This chapter offers speculative thoughts, asking what makes us perceive qualia as a “hard problem to solve.” David Chalmers calls this the “meta-problem of consciousness.” Solving it might make the qualia problem easier.

Scientists and philosophers disagree about qualia. Readers may dislike the author’s explanation or feel it misses the point, because he seeks to revise our existing conception of what needs explaining when we address qualia, rather than offer a solution to the qualia problem itself.

In the predictive-brain hypothesis, agents mentally model their own physiological changes, the external environment, and their continual interaction, predicting how changes in the world will affect future bodily and metabolic activity. Such agents might exhibit the sentience Jeremy Bentham proposed in An Introduction to the Principles of Morals and Legislation.

They see not merely a tree, but a place where soon-needed food will be found (connecting to Gibson’s affordance). They inhabit a temporally extended world (because they must predict) that is perceptually meaningful (because they must act adaptively). Bodily self-regulation, action, temporal extension, and so forth are necessary conditions for conscious experience.

These agents would surely behave as though sentient. But is that genuine sentience or merely its appearance? (A philosophical zombie?) This is not a good question to ask. The next step requires “predicting oneself as a predictor,” or Daniel Dennett’s “strange inversion.” A pub serves excellent Guinness and attracts many customers. Analysis reveals that the beer tastes better the sooner it is consumed after the keg is opened. The beer is good because the pub is busy, rather than the pub being busy because the beer is good. Dennett says the taste of candy (qualia) is similar: there are complex, subtle responses elicited by candy, and “the taste of candy” is merely their label.

Predicting agents construct simplified, shorthand self-models that anticipate these very complex responses and use them in everyday life. These shorthand self-models are qualia. Humans become absorbed in these models and suppose they contain something requiring scientific explanation.

Humans use visual perception to identify redness, large size, cuteness, dogginess, and other features and recognize a “labradoodle,” but do not consciously grasp the detailed processes identifying each feature. It is an efficient design to maintain only the simplified model needed to predict one’s state, without knowing all the internal workings.

Agents equipped with opaque, simplified self-models will find their subjective experiences strange and mysterious. Yet qualia are simply another tool for predicting themselves and other agents. Qualia exist, but our metaphysical puzzlement about them (the feeling that qualia require an explanation?) is mistaken.

Chapter 5. Expecting Better

Predictive processing is enormously useful in letting us understand what is happening with limited perceptual information, but it also creates bias. This biases primary perceptual experience itself, which seems objective and self-evident, rather than merely thought or judgment.

If I believe Jack is angry with me, I may read his neutral face as slightly angry and become more convinced he is angry, when my belief actually influenced my perception. Moreover, believing him angry changes my own expression and behavior, influencing him in turn and creating a vicious cycle.

Police mistaking phones for guns and shooting people is related (shooter bias, threat perception failure, affective realism effect). Even in places such as Philadelphia with equal Black and white populations, Black victims are far more numerous (80%). You will perceive what you feel.

Existing discrimination creates patterns; we learn those patterns and enact and reinforce discrimination. Removing discrimination in the real world is the ultimate goal, but takes time. Since humans learn so much through media, using novels, films, and mixed reality to create desirable fictional worlds can be very effective in challenging real-world prejudice and discrimination.

Better awareness of interoception, such as one’s heartbeat, also matters. Professor Sarah Garfinkel’s research finds that anxiety is common in people who 1) perceive internal signals inaccurately yet 2) are highly confident that their perceptions are accurate.

Our predictions are inevitably influenced by innate tendencies and past experience. We do not experience the world as it is. Nevertheless, several factors help keep our experiences from straying too far from reality.

First, the world itself continually corrects our mistaken predictions.
Second, physical and neural characteristics shared by our species make our experiences somewhat similar.
Third, humans are social animals who interact, share experiences, and adjust to one another.

The chapter’s conclusions:

Deeper understanding of predictive processing will allow us to reduce inappropriate biases through new training methods and so forth.
Most importantly, make the world better so inappropriate predictions (racial prejudices, for example) become less likely.
We can learn to predict better.

Chapter 6. Beyond the Naked Brain

From the extended cognition perspective, humans are natural-born cyborgs (one of the author’s earlier books is Natural-Born Cyborgs). This chapter discusses 1) epistemic action, contrasted with pragmatic action; 2) prediction’s control of action, including epistemic action; and 3) the predictive brain’s use of information wherever it is needed, inside or outside the brain.

Through cultural and evolutionary processes, humans have built environments and tools that complement the brain well, becoming increasingly intricately entangled with the world. Brain, body, pen, paper, and so forth combine into a new whole. (So far, a conventional introduction to extended cognition. See the author’s earlier Supersizing the Mind.)

Professor Carolyn Baum discovered that Alzheimer’s patients adapted to everyday life much better than expected. Visiting their homes revealed why: supportive environments were very well established, with notes and other aids. The environment performs some functions previously performed by the brain.

Skilled bartenders exploit differences in cocktail glass shapes, setting out the corresponding glasses as they take orders. The arranged glasses indicate order sequence and cocktail types, supporting memory. This is both epistemic and pragmatic action, since the glasses need to be taken out anyway.

In predictive processing, epistemic and pragmatic actions differ little: both reduce future prediction errors. Predictive agents will mix them appropriately at the right moments without specifically distinguishing the two.

Epistemic and pragmatic actions form loops. A soccer player uses epistemic action to predict where the ball will arrive, moves there at the right time, and thereby enables a pass. Expertise improves this ability. But increasing expertise relatively reduces the ability to notice unexpected events (such as a spectator running onto the pitch).

The North Sense device attaches to the chest and vibrates whenever the wearer turns north. After wearing it for a while, people naturally know the direction of home or their child’s school even from work, and constant directional awareness becomes normal. Removing it later causes considerable discomfort.

When the author and David Chalmers first wrote The Extended Mind in 1998, they could not explain the mechanism by which the brain uses suitable external resources at the right time. (That is, they could not explain how the principle of ecological assembly, the principle of ecological balance, and so forth are possible. —ak) He now thinks predictive processing is that mechanism. In the same situation, Otto consults his notebook and Inga recalls a memory: each is the most efficient available way of minimizing prediction errors. Both are epistemic actions, differing only in Otto’s being external and Inga’s internal.

One might distinguish a recruitment stage in which the brain draws in an appropriate external resource. After that stage, however, brain and external resource operate as one extended cognitive system. Note that recruitment must not require conscious effort for the whole to qualify as extended cognition.

Extended cognition also has ethical implications. Do we accept the athletic abilities of athletes with prosthetic limbs and the cognitive abilities of Alzheimer’s patients living in appropriately supportive environments as they are? Or recognize only the abilities of the biological body stripped of all artifacts?

Chapter 7. Hacking the Prediction Machine

Prediction and perception combine to form experience, so intervening in predictions can influence experience. Medical interventions and treatments can therefore use self-affirmation, talk therapy, deliberate placebos, VR, and other methods.

Prescribing placebos while wearing a neatly pressed doctor’s coat has more positive effects on conditions related to breathing, hormones, digestion, and other systems. Giving athletes a placebo said to increase red blood cell production improves performance. Placebo surgery (making an incision but doing nothing else) brings pain relief comparable to actual surgery.

Even honest placebos work. Drug-like packaging (despite being labeled “placebo”) and drug-like pills probably influence unconscious expectations.

When Parkinson’s patients receive apomorphine four times and a placebo on the fifth, the placebo produces effects similar to the drug. Starting with a placebo has no effect. Placebos work only after the body learns the actual drug’s physiological effects. Confidence in doctors, hospitals, drug packaging, rituals, and so on supports their effectiveness.

Showing severely burned patients a winter experience in VR reduced pain similarly to intravenous opioid analgesics. Brain scans showed reduced neuronal activity in pain-processing regions. Combining conventional treatment (analgesics) with VR also seems promising.

However, placebos cannot replace all medical interventions. They have no effect in many areas, and sometimes only treat symptoms. Much research is still needed for sound medical use of placebos, VR therapy, and similar approaches.

Self-affirmation can be as effective as placebos or rituals. After listing their positive personality traits and describing good things resulting from them, participants performed tasks such as problem solving much better than controls who wrote about others. Reported examples include elimination of gender gaps in spatial perception and mathematics after self-affirmation, a 62% reduction in academic achievement gaps affecting discriminated-against groups such as people of color, and a 40% reduction after a 15-minute self-affirmation exercise at school.

Experience can also be reframed. People feel fuller after the same food if deceived into thinking it contains more ingredients. The same milkshake feels more filling when labeled “rich and creamy” than “light and healthy.”

Over the past decade, research has increasingly explored psychedelics such as LSD for treating OCD, addiction, and other conditions. It is hypothesized that radically different experiences induced by psychedelics unsettle entrenched, problematic predictive models, creating room to update them.

Meditation quiets the ego similarly to psychedelics, with similar brain scan results. For example, meditation focused on breathing causes the predictive model to reduce weights assigned to other information and diminishes the influence of long-term predictions, allowing greater focus on the present. Attention control == prediction-weight control.

Conclusion. Ecologies of Prediction, Porous to the World

There is no raw perception. Every component generating experience, including perception, is active. Prediction errors help update current predictive models to fit incoming information. Predictive processing offers a new perspective integrating perception and action, but remains incomplete.

Although the interlude touches on it, predictive processing still cannot explain qualia. Nevertheless, the author believes it offers the most important clues toward doing so. Continued research from this perspective may solve the qualia problem.