Figure 3.

Markov blanket state separation in active inference. Middle: The internal (pink) and external states (white) of each agent are separated by a Markov blanket, comprised of sensory (grey) and active (green) states. This architecture can be applied to various forms of information processing, two of which are shown above and below. By associating the gradient flows of the Markov blanket partition with Bayesian belief updating, self-organization of internal states—in response to sensory fluctuations—can be thought of as perception, while active states couple internal states back to hidden external states vicariously, to provide a mathematical formulation of action and behavior. Top: Visual processing. Internal states are made up of the brain, which directs movement of the eye through abductor muscles as active states. Sensory states as the photoreceptors in the eye perceive the external states in the visual field of view. Bottom: Transcriptional machinery. Internal states are given by the gene expression levels plus epigenetic modifications. Intracellular components such as ribosomes, smooth and rough endoplasmic reticulum, and Golgi apparatus implement protein translation and delivery as the active states. Sensory states correspond to the surface states of the cell membrane, such as ligand receptors, ion channel states, and gap junctions between cells. External states are associated with extracellular concentrations and the states of other cells. Image credits: The cell schematic has been adapted from an image by Judith Stoffer supplied at the National Institute of General Medical Sciences, NIH (CC BY-NC 2.0). The brain schematic has been adapted from an image of a head cross-section on openclipart.org by Kevin David Pointon under a public domain license, and an image of an eye with abductor muscles from Wikimedia by Patrick J. Lynch (CC BY-NC 2.5).