A tale of two densities: active inference is enactive inference

Area of Science:

• Theoretical Neuroscience
• Computational Biology
• Philosophy of Mind

Background:

• The free-energy principle (FEP) and active inference are influential frameworks in neuroscience and biology.
• Existing literature often conflates FEP/active inference with related Bayesian brain theories like predictive processing.
• Misinterpretations of core constructs, such as generative models and variational densities, obscure their precise role.

Purpose of the Study:

• To clarify the interpretation of generative models and variational densities within the FEP and active inference.
• To address systematic misrepresentations in the literature stemming from conflation with other Bayesian frameworks.
• To propose and advocate for an enactive interpretation of active inference.

Main Methods:

• Critical analysis of the literature on the free-energy principle and active inference.
• Examination of two contrasting interpretations: structural representationalist vs. enactive.
• Theoretical argumentation for the enactive interpretation of generative and recognition models.

Main Results:

• Structural representationalist interpretations inadequately capture the function of generative and recognition models in active inference.
• Active inference under FEP involves belief-guided action selection (inference and control), not just representational mapping.
• An enactive interpretation, termed 'enactive inference,' better explains the role of these models in self-organizing systems.

Conclusions:

• Generative and recognition models in active inference are best understood as realizing inference and control.
• They facilitate self-organization and belief-guided action policies, rather than merely representing states.
• The proposed enactive inference framework offers a more accurate account of these processes in biological systems.