Markov blanket and sensory-active state partition

Friston, Wiese and Hobson (2020) argue that talking about a 'thing' requires a partition of all possible states into external, blanket and internal states.¹ They report that the term was coined by Pearl for Bayesian networks.¹ In their account, the contents of the blanket depend on the kind of graph. In a directed acyclic graph it holds the parents, children and parents of the children of a node. In an undirected graph it holds the neighbours. In a directed cyclic graph it holds only the parents.¹

Kirchhoff and colleagues (2018) divide the blanket into two kinds of state. Active states influence external states but are not influenced by them. Sensory states influence internal states but are not influenced by them.² Ramstead, Badcock and Friston (2018) state the same rule and add that internal states only 'see' external states through the blanket.³ Pio-Lopez and colleagues (2022) repeat it in work on morphogenesis, writing the blanket as the product of sensory and active states.⁹ Allen and Friston (2018) describe the loop that results: sensory states depend on hidden states, which depend on active states, which depend on internal states.⁵

Kirchhoff and colleagues (2018) state that, as shown by Friston, the conditional independencies of a blanket induce active inference. They describe this as the tendency of random dynamical systems to minimize free energy on average.⁴ They read the result as self-evidencing dynamics behind the autonomous organization of life, and they associate the internal states with a model of the external states.⁴ Ramstead and colleagues (2018) use the blanket to separate system from environment, which they need to tell surprise apart from free energy.³ Krakauer and colleagues (2020) say the free-energy principle, like their theory, starts from first principles following Schrödinger. They say adaptive systems minimize free energy by constructing partitions that separate the organism from its environment.⁸

Palacios and colleagues (2020) ask how the concept extends to hierarchical systems. They argue that a macroscopic blanket must be composed of microscopic blankets, so that blanket formation at one level is tied to blankets 'all the way down'.⁶ They call self-organisation a recursive process of boundary formation and offer a simulation as proof of concept.⁶ Ramstead and colleagues (2018) describe how to build the next level. States at one level are partitioned into internal subsets and their blankets, and the next level takes mixtures of the blanket states. Internal states can be ignored in that step.⁷