Interaction

Markov blankets and the separation of internal and external states

  • In computer science, a Markov blanket of an element α is the set of nodes that shields α from influence by all other elements in a dynamical system (Pearl, 2014).

  • Let A denote a collection of nodes representing a dynamical system; a Markov blanket for A comprises:

    • a) ALL the nodes that can influence ANY member of A (parents),

    • b) ALL the nodes that ANY member of A can influence (children),

    • c) ALL the nodes that can influence the children of ANY member of A (spouses), and

    • d) NO other nodes besides these.

  • The question posed: identify the Markov blanket of node X in terms of its parent, children, and spouse nodes.

  • This framework underpins how systems infer states and respond to stimuli while being shielded from irrelevant parts of the world.

Markov blankets and the Markov blanket paradox

  • How can living organisms adapt to a world that is, by definition, cut off from them? This tension is dubbed the Markov blanket paradox.

  • Separation and mediation by membranes:

    • The cell membrane acts as a Markov blanket separating internal states (cytoplasm, organelles, DNA) from the external environment.

    • It mediates exchange of matter, energy, and information (nutrients, wastes, ions) between cell and environment.

    • As a physical boundary, it creates a statistical partition between intracellular and extracellular states.

    • As a communication medium, it allows interactions with the environment.

  • Living cells access their internal states directly (limited by the membrane) but respond to environmental stimuli, enabling adaptation and survival.

  • Key terms: Separation; Mediation; and the idea that life persists by operating behind a Markov blanket.

E. coli: a paradigmatic Markov-blanket-bound organism

  • Organism: Escherichia coli (E. coli) is a prokaryotic, single-celled bacterium; its outer membrane functions as a Markov blanket, separating it from its environment.

  • Chemotaxis: E. coli can move in response to chemical gradients (up or down concentration of attractants/repellants).

    • Attractants include nutrients such as glucose; repellants include acids and toxins.

    • Chemoreceptors enable detection of changes in chemical concentrations.

  • Movement behavior: E. coli exhibits a run-and-tumble random walk.

    • In a gradient, run durations become longer in the direction of increasing attractant (biased random walk).

    • If attractant concentration increases (gradient present), runs lengthen in the positive direction; if not, the bacterium tends to tumble more to reorient.

  • Scenario A (no gradient): smooth runs punctuated by short tumbles; overall random exploration.

  • Scenario B (gradient present): longer runs toward higher attractant concentration; biased movement up-gradient.

Running and tumbling: motor behavior in E. coli

  • Running: flagella rotate anti-clockwise, bundle together, propel forward.

  • Tumbling: if attractant gradient weakens or plateaus, motor rotation changes to clockwise, flagella unbundle, reorient randomly.

  • The run-and-tumble strategy enables effective foraging in fluctuating environments.

  • The running/tumbling dynamics are a concrete example of how a biological system minimizes prediction error by acting to reduce surprise in its environment.

Prediction error minimization (PEM) and active inference in biology

  • PEM is a framework in which organisms aim to minimize surprise, prediction error, or uncertainty while operating behind a Markov blanket using internal models.

  • If predicted outcomes diverge from actual outcomes, organisms can minimize free energy by:

    • acting on the world to change sensory input, or

    • updating internal models to improve predictive accuracy.

  • Core figures: Karl Friston, Jakob Hohwy, and collaborators have developed PEM as a unifying account of perception, action, and learning.

  • Key idea: life maintains structure by reducing discrepancy between internal predictions and incoming sensory data.

Quantifying mismatch: Accuracy and Divergence in E. coli chemotaxis

  • When there is a mismatch between expected and actual sensory input, the discrepancy is termed surprise, prediction error, or free energy.

  • If the internal model predicts increasing attractant concentration, the organism selects actions that maximize model accuracy:

    • α  argmaxα(Accuracy)\alpha \; \arg\max_{\alpha}(\text{Accuracy})

  • If there is a mismatch (gradient does not rise as predicted), the organism updates its beliefs to minimize divergence:

    • μ=Divergence\mu = \text{Divergence}

  • The objective is to minimize prediction error via action (alterting sensory input) or belief updating (adjusting priors).

  • The PEM view provides a dynamic account of how a simple, non-neural organism like E. coli can exhibit goal-directed, adaptive behavior.

The organism–environment loop under PEM

  • Organisms operate behind a Markov blanket, using internal models to predict sensory input.

  • They act to confirm predictions or update internal models when there is prediction error.

  • Benefits of minimizing free energy and prediction error: resist entropy, minimize surprise and uncertainty, maintain structural and functional integrity, persist in states aligned with prior preferences over expected sensory states.

  • Prior preferences encode survival-related goals (e.g., energy acquisition, threat avoidance, reproduction).

Biochemical mechanism in E. coli: Perception, internal model, and action

  • Perception: E. coli samples chemical gradients over time via chemoreceptors; perceives gradient direction and magnitude.

  • Internal model: receptor–response mapping translates changes in attractant/repellent gradients into movement commands.

    • ↑ attractant gradient → keep swimming (run);

    • ↓ attractant gradient → change direction (tumble);

    • ↓ repellant gradient → keep swimming (run);

    • ↑ repellant gradient → change direction (tumble);

    • No gradient changes → unbiased random walk.

  • Action: movement described as biased/unbiased random walk (run/tumble) depending on gradient information.

  • Prediction error minimization: the running/tumbling cycle encodes continuous updating of the internal model to align predictions with sensory data.

Active inference and E. coli (no nervous system required)

  • Active inference framework applies to E. coli: actions are selected to minimize expected surprise (free energy) even in simple organisms.

  • The claim: PEM does not require consciousness or a nervous system; it operates via biochemical signaling and receptor dynamics.

  • Simulation and visualization: a variety of computational studies illustrate chemotaxis under PEM-like dynamics.

Phototropic plants and PEM

  • Plants lack a nervous system but can perform active inference-like processing: phototropism is the directed growth toward light.

  • Phototropins detect light direction and intensity, enabling a gradient to be perceived.

  • Prediction error arises when light gradient deviates from the plant’s expectation; auxin redistribution acts to correct the gradient by promoting growth on the shaded side.

  • Result: stem exhibits positive phototropism (bends toward light) and roots exhibit negative phototropism (grow away from light).

  • Historical anchors: von Sachs, Darwin & Darwin, Cholodny, Went; traditional view held that plants are passive, but PEM offers a biochemical mechanism for adaptive orientation.

Biomats (microbial mats) and layered metabolism

  • Biomats or microbial mats are multi-layered sheets or biofilms of microbes; among the oldest biosystems.

  • Fossil and molecular evidence places ancient biomats at ~3.4–3.7 billion years ago; stromatolites are key structures formed by layered microbial mats.

  • Biomats have existed for roughly 3.96 billion years, predating humans by tens of billions of years in evolutionary terms.

  • Biomats form when cyanobacteria link to create filamentous networks; EPS (extracellular polymeric substances) glue cells into a cohesive matrix.

  • Layered structure of modern conceptual biomats:

    • Top layer: oxygen-rich, well-lit; cyanobacteria perform oxygenic photosynthesis: 6CO<em>2+6H</em>2O+sunlightC<em>6H</em>12O<em>6+6O</em>26\mathrm{CO}<em>2 + 6\mathrm{H}</em>2\mathrm{O} + \text{sunlight} \rightarrow \mathrm{C}<em>6\mathrm{H}</em>{12}\mathrm{O}<em>6 + 6\mathrm{O}</em>2

    • Middle layer: purple bacteria perform anoxygenic photosynthesis: CO<em>2+2H</em>2S+sunlightcarbohydrates+2S+H2O\mathrm{CO}<em>2 + 2\mathrm{H}</em>2\mathrm{S} + \text{sunlight} \rightarrow \text{carbohydrates} + 2\mathrm{S} + \mathrm{H}_2\mathrm{O}

    • Deeper layers: heterotrophic bacteria decompose organic matter, releasing CO2 and other compounds; oxygen and sulfide cycles are coupled between layers.

  • The mats circulate nutrients and enable energy flow in microbial ecosystems.

Quorum sensing in biomats

  • Quorum sensing is a cell-to-cell communication mechanism that enables bacteria to sense cell density via signaling molecules called autoinducers.

  • Autoinducer concentration acts as a proxy for cell density; regulatory effects switch on/off genes when thresholds are crossed.

  • Phase-dependent changes:

    • Low density: autoinducer levels below threshold; gene regulation remains off; microbes conserve energy.

    • High density: autoinducers accumulate; genes regulated by quorum sensing are switched on, coordinating behaviors such as EPS secretion.

  • This coordinated behavior exemplifies a collective, PEM-like dynamic: individuals align actions based on noisy sensory inputs from their community and environment.

Perception, internal models, and action in biomats (quorum sensing view)

  • Perception: each microbe detects autoinducer concentrations.

  • Internal model: receptor–response mappings translate signal levels into regulatory states.

  • Action: microbes adjust behaviors (EPS secretion, enzyme production, toxin production) in response to signal levels.

  • Public goods: EPS, enzymes, and toxins are communal resources that benefit the entire mat but cost energy/resources to produce.

  • Social cheaters: organisms that exploit public goods without contributing investments, gaining fitness advantages under certain conditions.

  • Example: Pseudomonas aeruginosa can accumulate social cheaters after ~100 generations under particular conditions, threatening mat stability.

  • Cheating can destabilize cooperation; multiple strategies exist to curb cheating and sustain mat integrity.

Strategies to counter cheaters in biomats

  • STRATEGY 1: Toxin-mediated policing

    • Some microbes produce toxins targeting cheaters; effective when toxins are powerful, durable, and inexpensive to produce.

  • STRATEGY 2: Kin selection

    • Clonal growth means cooperating benefits relatives (copies of oneself), making cheating less favorable.

  • STRATEGY 3: Spatial structure

    • Limited diffusion of public goods means benefits are localized to nearby producers; distant cheaters gain less advantage.

  • These strategies illustrate how evolutionary and ecological principles interface with PEM-like dynamics to maintain cooperative behavior.

Survival as a guiding preference under PEM

  • Survival is a fundamental preference encoded in an organism’s prior expectations.

  • PEM suggests that resisting entropy, minimizing surprise, and maintaining structural integrity support persistence in states aligned with prior preferences.

  • Comparative summary:

    • E. coli: perception of chemical gradients; internal models predict attractant concentration; action via run/tumble to exploit gradients.

    • Microbial mats: perception of autoinducer levels; internal models regulate EPS, enzymes, and toxins to coordinate community behavior.

    • Phototropic plants: perception of light gradients; internal models predict light direction; actions via auxin redistribution leading to bending toward light.

PEM extended to human intelligence and everyday cognition

  • The PEM framework can be extended to human cognition and behavior, including everyday decisions and motor tasks (e.g., cycling).

  • Example: a cyclist encountering a pedestrian or rain collision risks involves mismatch between predicted and actual sensory input; the cyclist may act (brake) to minimize prediction error or update beliefs about rain likelihood.

  • Predictions and updates can be framed as:

    • Accuracy-driven actions to maximize alignment with expectations, or

    • Divergence-driven belief updates to reduce future surprise.

PEM and everyday examples beyond the microcosm

  • Example: rain on a path and tyre grip:

    • Internal model expects dry conditions; sensory input (slippery tyres) mismatches the expectation.

    • Update beliefs about the path to minimize divergence; even after rain stops, residual moisture may keep the path uncertain.

  • The broader idea: PEM provides a unified lens for how organisms and agents navigate uncertain environments by balancing action and belief updates.

Theory of everything: PEM as a unifying principle (Hipólito 2019)

  • Hipólito (2019) argues for PEM as a simple, universal theory that can be applied to nearly everything from particles to minds.

  • Core claim: self-organizing systems that can be described via Markov blankets fall under PEM as a unifying principle.

  • Examples of systems described by Markov blankets and PEM:

    • Stars, planets, environments, organisms, brains, neural nets, neurons, and neuronal processes.

  • The idea is to derive complex natural phenomena from a small set of principles about Markov blankets and prediction error minimization.

Theory of everything: Markov blankets and system states

  • An organism or a thing is composed of Markov blankets of smaller things.

  • A system state is a member of a set X: x = {η, s, a, μ} ∈ X, where:

    • Internal states (η)

    • Sensory states (s)

    • Active states (a)

    • External states (μ)

  • Communication pattern:

    • External states cause changes in internal states via sensory states.

    • Internal states couple back to external states via active states.

  • Each Markov blanket partitions internal vs external states; a further partition between sensory and active states can be considered.

Self-entropy, risk, and ambiguity in PEM

  • Self-entropy measures how unpredictable a system is.

  • Types of uncertainty:

    • Meaning: risk or complexity (uncertainty about the outside world).

    • Ambiguity: uncertainty about what a system senses.

  • Divergence: discrepancy between prior and posterior distributions over external states.

  • Self-organizing systems aim to minimize self-entropy by reducing risk and improving reliability of senses (reducing ambiguity), while maintaining beneficial interactions with the environment.

  • Question posed: Do you agree with Hipólito that PEM is a simple theory of everything? (Inês Hipólito)

PEM and good writing: leveraging PEM to structure text

  • Imagine an intelligent reader who is predicting what comes next in your text.

  • Good writing aims to minimize prediction error by providing:

    • Clear, coherent logical structure and reasoning;

    • Relevance and precision in language; accessibility of information;

    • Predictable flow that allows the reader to anticipate and follow arguments.

  • Cont’d: Introducing surprise strategically by providing novel or nuanced information to encourage updating internal models and reducing divergence.

How good writing aligns with PEM principles

  • Writing should persist in states consistent with prior preferences over expected sensory states (the reader’s expectations):

    • Convincing the reader of your position with reasons;

    • Demonstrating understanding and engagement with the subject matter;

    • Aiming for high-quality assessment outcomes (A to A+ on assignments).

Appendix: Python code for simulating E. coli chemotaxis

  • The appendix provides Python code to simulate the chemotactic behavior of a single bacterium in a 2D field with a central attractant.

  • Key components:

    • import numpy as np; import matplotlib.pyplot as plt; animation utilities.

    • Parameters:

    • gridsize = 50; steps = 300; attractantcenter = (25, 25);

    • attractantstrength = 100; decay = 0.05; stepsize = 0.7;

    • Attractant field function:

    • r=(xx<em>c)2+(yy</em>c)2r = \sqrt{(x- x<em>c)^2 + (y- y</em>c)^2}

    • attractant ext field(pos)=attractant ext strengthedecayr\text{attractant ext{ }field}(pos) = \text{attractant ext{ }strength} \cdot e^{-\text{decay} \cdot r}

    • Precompute attractant grid for visualization (Gaussian-like field) and create a meshgrid of X, Y with Z representing the attractant concentration at each point.

    • Initialize bacterium position pos randomly in the grid; initial angle random.

    • Simulation loop (for each step):

    • deltapos = stepsize * [cos(angle), sin(angle)];

    • newpos = pos + deltapos; clamp to [0, grid_size];

    • newconc = attractantfield(new_pos);

    • delta = newconc - prevconc;

    • If delta > 0: continue; if delta < 0: rotate angle randomly; if delta == 0: randomly decide to rotate or not;

    • Update pos and prev_conc; store path history.

    • Visualization: creates an animation of the bacterium path on top of the attractant field; marks attractant center with a green dot; colorbar shows attractant concentration.

  • The code illustrates a PEM-like chemotaxis demonstration where the organism follows concentration gradients via simple rules tied to local sensing and motor changes.

Summary: PEM framework across biological scales and human cognition

  • PEM and Markov blankets provide a unifying lens to view adaptation across organisms, microbial communities, plants, and potentially human cognition.

  • Key themes:

    • Internal models predict sensory input; actions serve to minimize prediction error or to update models when errors occur.

    • Separation via Markov blankets allows organisms to persist while interacting with their environment.

    • Cooperation in biomats emerges from local interactions (quorum sensing) and ecological strategies to deter cheating.

    • Writing and communication can be viewed as a PEM exercise: effective writing reduces prediction error and strategically introduces novel information to update the reader’s model.

  • The synthesis aims to show how simple, local rules and interactions can give rise to complex, adaptive behavior and even broad philosophical claims about information, prediction, and life itself.