Predictive processing
Explanation
A traditional picture of perception: the senses gather data from the world and the brain processes them bottom-up to form representations. Predictive processing completely inverts this picture. According to authors such as Karl Friston, Andy Clark and Jakob Hohwy, the brain is fundamentally a prediction machine: it constantly generates hypotheses about which stimuli it will receive, and only processes in detail the discrepancies between prediction and input.
The proposed architecture is hierarchical. The upper cortical layers generate top-down predictions about what the lower layers should be detecting. The lower layers compute the prediction error (the difference between what is predicted and what is sensed) and transmit that error upward. The whole system adjusts by minimising errors, updating its internal models when predictions fail.
This architecture explains a very everyday phenomenon: most of perception is "anticipated" and invisible. We do not notice background noise, the feel of clothes, the chafing of a shoe, because our brain already predicts them and their errors are minimal. But if something changes suddenly (an unexpected noise, a strange touch), the prediction error spikes and the stimulus becomes conscious.
Applied to consciousness, the theory suggests that subjective experience is the best hypothetical model the brain has built about the causes of its sensory inputs, including the causes it itself generates (the body's interoceptive signals). The reality we experience is not what is "out there" but the brain's best conjecture, continuously corrected by prediction errors.
Anil Seth has popularised an evocative consequence: "perception is controlled hallucination". Under normal conditions, the external world controls our predictions by providing useful errors. Under extreme conditions (sensory deprivation, psychedelics, schizophrenia) predictions can dominate without external control, and what we live becomes decoupled from reality. Dreams would be uncorrected hallucinations.
Predictive processing has transformed cognitive neuroscience, clinical psychology (explaining psychosis as an error in the confidence assigned to predictions), learning and artificial intelligence. It is one of the most fertile and unifying theories of recent decades, and is starting to provide a common mathematical framework for understanding perception, action, learning and, possibly, consciousness.
Strengths
- Powerful unifying framework: perception, action, learning, emotion.
- Computationally articulated and empirically fertile.
- Connections with clinical neuroscience (psychosis, autism).
- Naturally integrates enaction and the body.
Main critiques
- The free energy principle is so general that it seems unfalsifiable.
- Does not directly address the hard problem of consciousness.
- Risk of empty mathematisation if not grounded neurally.
- Some variants are accused of accounting for everything and therefore for nothing.