Predictive processing and interoceptive inference

Calvo and Baluška (2015) applied the idea outside animals. Their working hypothesis was that adaptive plant behaviour can only take place through a mechanism that predicts sensory states¹. They separated weaker and stronger readings of anticipation. On the weaker reading, a system models the sources of its stimulation internally. Stronger forms of anticipation that explain away internal modelling, as discussed by Stepp and Turvey, could not be ruled out beforehand¹. The authors treated predictive coding and strong anticipation as working hypotheses to be tested empirically, not as settled accounts¹.

Constant and colleagues (2018) defined a generative model as a mapping from sensory observations and action policies to external, hidden causes⁵. Ramstead, Badcock and Friston (2018) described two routes to reducing prediction error. An organism can improve its predictions by altering internal states, which they call perception. Or it can act on the world to confirm its predictions, which they call action⁷. Constant and colleagues also derived a special case. Under identical synchronisation, the agent's model of its environment is exactly the same as the environment's model of the agent⁶.

Allen and Friston (2018) argued that predictive coding in comparator models of motor or social control is clearly cognitivist. In those models, rich internal models of a hidden world do the work of cognition³. They noted that Clark presents predictive processing as the most radical step away from a passive view of neural processing, while Hohwy has argued that such approaches are inherently disembodied³. They also listed critiques of these accounts. Critics charge that they are circular, unfalsifiable, or post-hoc explanations. The authors then asked whether enactivism could answer these charges⁴.

Friston, Levin, Sengupta and Pezzulo (2015) simulated self-assembly by specifying a generative model and the dynamics of the environment. The equations of motion for the system and its environment then follow². Pio-Lopez and colleagues (2022) describe cells whose internal states minimise precision-weighted prediction error. The predictions concern signalling states sensed from neighbouring cells, and the authors identify this scheme with predictive coding⁸. They note that earlier simulations left the generative model and its precision untouched. Kuchling and colleagues altered external parameters to produce disorders of morphogenesis, and Friston and colleagues altered extracellular signals⁹.