Perception as active inference
Explanation
Perception as active inference is a proposal that has developed especially since the late twentieth century and has roots in Helmholtz (nineteenth century), who already thought of perception as unconscious inference. The modern idea, associated with names such as Karl Friston, Andy Clark, Jakob Hohwy, holds that the brain does not passively receive sensory information, but continuously generates predictions about the world and compares those predictions with the incoming sensory information. Perception is the result of that comparison.
The brain, according to this model, is essentially a hierarchical prediction machine. Higher levels generate generative models about the probable causes of sensory data; lower levels compare those predictions with the actual input. When there is a match, perception is stable. When there is a prediction error, that error is propagated upwards to update the models. Conscious perception would be the brain's best hypothesis about what is producing the sensory data at each moment.
Friston's active inference extends this idea: we do not only predict passively; we also act so that the world matches our predictions. If I predict I am safe and the sensory information suggests danger, I can update the model or act (flee, defend myself) so that reality confirms the desirable model. Perception and action are two sides of the same coin: minimisation of expected surprise, or free energy, in Friston's terms.
For the theory of consciousness, this framework is highly productive. It explains classical perceptual phenomena (illusions, bistable perceptions, hallucinations), allows modelling of psychiatric disorders (psychosis as excess of un-inhibited prediction error, autism as difficulty in constructing flexible generative models), and connects with theories such as REBUS on psychedelics (psychedelics flatten the predictive hierarchies, allowing normally inhibited contents to emerge).
The connection with consciousness is established in several ways. Perceptual consciousness would be the content of the brain's hypotheses in their best explanation of the data. Self-consciousness would be a special case: the brain generates a model of itself as a predictive agent and experiences that model as I. Unusual experiences (dreams, hallucinations, depersonalisation) correspond to specific configurations of the predictive system, where certain balances between prediction and error fail.
Criticisms and debates are several. Some question whether the principle of free-energy minimisation is specific enough to predict empirically, or whether it is a very general formulation that applies to everything. Others note that active inference explains functional cognition well but leaves open the hard problem of qualia. Defenders reply that it is the most unifying framework we have today for integrating perception, action, learning and emotion, and that its explanatory power justifies its dominant status in contemporary computational neuroscience.
Strengths
- Integrates centuries of tradition (Helmholtz) with contemporary formalisms.
- Explains illusions, hallucinations and expectations.
- Compatible with hierarchical neuroscience.
- Foundation of predictive processing and the free energy principle.
Main critiques
- The formalism (free energy) can be vacuous if not constrained.
- Possibility of explaining anything reduces predictive power.
- Does not directly address the hard problem.