Dynamic Markov Blanket Detection

• Dynamic Markov blanket detection is a technique that identifies and tracks minimal variable subsets which shield dependencies in complex stochastic models.
• It leverages context-specific independence tests via finite-state automata in Tsetlin Machines and Bayesian EM procedures in state-space models to adaptively prune irrelevant variables.
• Empirical studies on synthetic and physics-based systems show that this approach enhances interpretability and computational efficiency in modeling dynamic processes.

Dynamic Markov blanket detection refers to the identification and tracking of variable subsets (Markov blankets) that optimally shield or mediate the dependencies between subsystems within complex stochastic models, allowing for minimal yet sufficient representations of causal or informational structure that may change over time, context, or system state. Recent developments span both symbolic machine learning (Tsetlin Machine with Markov boundary-guided pruning) and generative modeling (Free Energy Principle-based unsupervised object discovery), linking the concept to advances in structure learning, Bayesian inference, and unsupervised macroscopic physics modeling (Granmo et al., 2023, Beck et al., 28 Feb 2025).

1. Formal Definition and Conceptual Basis

A Markov blanket for a target random variable $Y$ in a set $\mathcal{X} = \{X_1, \ldots, X_n\}$ is a subset $S \subseteq \mathcal{X}$ such that $Y$ is conditionally independent of all other variables given $S$, i.e.,

$Y \perp\!\!\!\perp \mathcal{X}\setminus S \mid S.$

If $S$ is minimal with respect to this property (no proper subset of $S$ is itself a blanket), it is termed a Markov boundary (Granmo et al., 2023).

Under the Free Energy Principle (FEP), systems are partitioned into internal ($z$), blanket ($b$), and external ($s$) variables. The blanket $b$ provides the only interface: $p(s,z|b) = p(s|b) p(z|b)$. In dynamical systems, this partition corresponds to Langevin-like dynamics where direct coupling between $s$ and $z$ is mediated via $b$, ensuring $z_\tau \perp s_\tau | b_\tau$ along any trajectory (Beck et al., 28 Feb 2025).

Dynamic Markov blanket detection extends these notions to contexts where the blanket structure may vary across samples, system configurations, or time, and addresses the adaptive identification of such structures via data-driven algorithms.

2. Markov Blanket Detection in Tsetlin Machines

The Markov blanket concept is operationalized in the Tsetlin Machine (TM) through the design of the Context-Specific Independence Automaton (CS-IA), as introduced in "Learning Minimalistic Tsetlin Machine Clauses with Markov Boundary-Guided Pruning" (Granmo et al., 2023).

Each clause and literal in the TM is assigned a finite-state automaton, the CS-IA, tasked with deciding—based on streaming data—whether to prune (“Prune” action) or retain (“Keep” action) a literal.

CS-IA Mechanism

• States: $1,\ldots, 2N$; $1,\ldots,N$ correspond to "Prune", $N+1,\ldots,2N$ to "Keep".
• Initialization: Strong keep bias ($s_i(0) = 2N$).
• Feedback Scenarios:

1. Full clause: $C(\mathbf X) = 1$ with $X_i$ included.
2. Reduced clause: $C_{-i}(\mathbf X_{-i}) = 1$ simulating $X_i$ excluded.

• Transition Probabilities depend on outcomes (target $Y$, presence of $X_i$). The parameter $d > 1$ controls conservativeness.

Literal pruning is triggered when the state crosses $N$, systematically eliminating empirically independent variables and converging to context-specific Markov boundaries.

TM Training Workflow

After each learning epoch:

1. Evaluate clause outputs.
2. Apply standard (Type I/II) TM feedback.
3. For each literal:
   • Update CS-IA using full/reduced clause scenarios.
   • Prune if CS-IA indicates independence.

This process allows clauses to dynamically shed superfluous literals, yielding parsimonious, interpretable feature sets matching Markov boundaries.

3. Dynamic Markov Blanket Discovery in Stochastic Dynamical Systems

In the context of the Free Energy Principle and physics-based generative modeling, dynamic Markov blanket detection is implemented as a Bayesian EM procedure over state-space models with explicit roles for internal, blanket, and external states (Beck et al., 28 Feb 2025).

Model Structure

• Observations: Microscopic trajectories $\{y_i(t) \in \mathbb{R}^D\}$.
• Latent Variables:
  • Continuous macroscopic states: $s_t$, $b_t$, $z_t$.
  • Discrete labels: $\omega_i(t) \in \{S, B, Z\}$, i.e., external, blanket, or internal assignment.
• Process: Time-evolving linear-Gaussian state-space for $(s_t, b_t, z_t)$; $\omega_i(t)$-labels as Markov chains constrained so $S\leftrightarrow Z$ transitions are forbidden.
• Joint Distribution:

$$p(x_{1:T}, \omega_{1:T}, s_{1:T},b_{1:T},z_{1:T}|\theta)$$

factors to reflect conditional independence and blanket structure.

Algorithmic Summary

1. E-step:
   • Update $q_\omega$ (assignment labels) via forward-backward algorithm (HMM).
   • Update $q_{sbz}$ (macrostates) via Kalman smoothing.
2. M-step:
   • Closed-form maximization of $$A, B, \Sigma_{sbz}, \{C^r, D^r, \Sigma^r\}, T(\cdot)$$.
3. Iterate: Until convergence in ELBO.

The dynamic assignment $\omega_i(t)$ allows for elements to shift roles (internal ↔ blanket ↔ external) as macroscopic objects move, exchange matter, or undergo phase transitions.

4. Theoretical Analysis and Examples

TM Convergence Guarantees

For a toy Bayesian network with $Y$ and potential Markov blanket variables $X_1, X_2$, analysis shows:

• $X_1$ is retained in the clause iff

$$[P(Y=1|X_1=1) - P(Y=0|X_1=1)] - [P(Y=1|X_2=1) - P(Y=0|X_2=1)] > d,$$

for conservativeness $d$.

• $X_2$ is pruned if $d > 0$.

Thus, with appropriate hyperparameter settings and infinite data, TM clauses almost surely converge to true Markov boundaries (Granmo et al., 2023).

Dynamic Label Evolution

In the FEP-based method, label vectors $p_i(t) = (p_i^S, p_i^B, p_i^Z)^T$ evolve via a time-inhomogeneous Markov chain governed by $T(b_t)$, enabling responsive, contextually accurate assignments to boundary roles as $b_t$ migrates or the object's structure evolves (Beck et al., 28 Feb 2025).

5. Algorithmic and Computational Considerations

TM with CS-IA

The TM Type III feedback process (integration of CS-IA at the clause and literal level) introduces negligible additional overhead relative to base TM updates. Pruning decisions are made online with finite-state automata, scaling linearly in the number of literals and clauses.

DMBD for Physics Discovery

The total computational cost per EM iteration is dominated by $O(N |\mathcal{S}|^2 T)$ for the HMM label updates and $O((d_s + d_b + d_z)^3 T)$ for Kalman smoothing on macrostates ($|\mathcal{S}| = 3$) (Beck et al., 28 Feb 2025).

6. Empirical Results and Applications

Tsetlin Machine Context

Empirical studies on synthetic data with known Markov boundaries demonstrate that TM clauses equipped with CS-IA reliably prune down to minimal informative features, preserving predictive accuracy while improving interpretability and sparseness of the learned representation (Granmo et al., 2023).

FEP/DMBD Applications

Dynamic Markov blanket detection has been demonstrated on physical and synthetic systems:

• Newton’s cradle: Correct identification of objects and transient boundary roles for collision balls.
• Burning fuse: Precise tracking of the reaction front as a moving blanket.
• Lorenz attractor: Differentiation of phase-space lobes, with transitions labeled as blanket boundaries.
• Synthetic cell models: Emergence of interpretable nucleus, membrane, and environmental compartments.

In all cases, detected boundaries aligned with intuitive decompositions, and regressed macroscopic dynamics accurately recapitulated low-dimensional system laws (Beck et al., 28 Feb 2025).

7. Limitations and Directions for Extension

Identified challenges and future directions include:

• Hyperparameter selection for the conservative/pruning tradeoff ($d$ in TM-CS-IA).
• Extension to multivalued or continuous-variable settings, which require alternative automaton or probabilistic architectures.
• Integration with constraint-based Bayesian network structure search and richer context-specific independence tests.
• The potential for hybrid approaches between symbolic structure learning and generative modeling frameworks.

These advances suggest promising avenues for joint inference of minimal sufficient feature sets and interpretable macroscopic laws in high-dimensional and temporally evolving systems.