A joint distribution is an answer, not a practical memory format

Pearl begins from a cognitive and computational dissatisfaction. Human reasoning integrates uncertain reports, symptoms, expectations and background knowledge, yet people do not appear to consult one enormous table containing a probability for every conjunction of every proposition. Such a table is mathematically complete but psychologically implausible and computationally explosive. With n binary variables it requires values for 2ⁿ assignments before constraints reduce the count. A representation can be valid and still be the wrong unit of construction.

The paper proposes that the elementary units of judgement are lower-order relations among small clusters of semantically related propositions. A clinician may hesitate to quantify an entire multivariate distribution but confidently say that a complication depends on a disease once age is held fixed, or that two findings become irrelevant after their common cause is known. Conditional-dependence judgements can be more accessible than a complete numerical specification. Pearl therefore gives graph structure a role prior to bulk arithmetic.

A Bayesian network is a directed acyclic graph. Nodes stand for propositions or variables. In Pearl’s 1985 presentation, arcs signify direct causal influences; a conditional probability table measures the strength associated with each child-parent relation. If Sᵢ is the parent set of Xᵢ, the local assessments multiply into a joint model. The graph is not merely a picture pasted on top of a probability table. It is a recipe for constructing the table and a map of the independencies that keep it compact.

The product has two directions of interpretation. Generatively, start at root variables and sample each child from its conditional distribution given its parents. Algebraically, multiply the local factors to obtain a normalized joint distribution that supports every ordinary probability query. Elicitation therefore remains local while semantics remain global. If the graph and conditionals are coherent, there is no need to fill the joint table by hand.

This combination answers an old tension between symbolic and numerical knowledge. A purely symbolic edge can say that one proposition bears on another but not how strongly. An unstructured numerical array can answer probability questions but conceal why particular entries belong together. The network keeps the qualitative skeleton and quantitative assessment aligned. Change a conditional table and the strength changes without redrawing the domain; change the graph and the independence commitments change even before new numbers are supplied.

It also separates direct assessment from derived belief. Experts provide a root prior or a family conditional, not every posterior they might later be asked to report. Once evidence arrives, the posterior is an implication of the assembled model. That discipline prevents a system from storing mutually inconsistent answers to overlapping questions. It does not guarantee correct assessments, but it gives disagreement a location: structure, parameter, or observation can be challenged separately.

Compression is earned by independence, not supplied by drawing arrows. A complete DAG can represent any joint distribution but saves little. Sparse parent sets assert that once a node’s direct parents are known, more remote non-descendants add no information about it. Those assertions can be false. A Bayesian network is economical because it commits to a structure; the commitment must be defended by knowledge or tested where possible.

A path can carry influence, block it, or open because of evidence

A graph earns its keep by making relevance visible. Pearl’s paper describes a separation criterion: if every path between two variables is blocked by a conditioning set, the variables are independent given that set. Modern readers call the directed version d-separation. The crucial point is that blocking depends on the local shape of the path. Chains, forks and head-to-head meetings do not behave alike.

In a chain A→B→C, observing B blocks information between A and C. Once the mediator is known, learning the upstream cause does not change the distribution of the downstream variable under the model. In a fork A←B→C, observing the common parent B likewise blocks association between its children. Without B, one child is evidence about the other because it carries information about their shared source.

A collider reverses the rule. In A→B←C, the path is blocked when B and its descendants are unobserved: two independent causes do not become associated merely because they can produce the same effect. Observe B, however, and the causes become dependent. If the shared effect occurred, evidence against A raises the relative plausibility of C. Observing a descendant of B can also open the path because it informs B. Pearl’s Figure 1 explicitly illustrates this special head-to-head treatment.

This is more than a test for zero correlation. Separation describes an independence that holds for every distribution compatible with the graph’s factorization. Numerical coincidences may produce extra independencies, but the graph guarantees only its structural ones. Conversely, an open path does not prove a strong association; it says the structure permits information to pass.

Graph reading also prevents a common error about direction. An arrow C→F helps define the factorization P(C)P(F|C), yet evidence about F can update belief in C through Bayes’ rule. The modelling direction and the information-flow direction are different concepts. Diagnostic reasoning routinely moves from effect to cause even when the causal model points from cause to effect.

The network is meant to compute from inside itself

Pearl’s title matters. The network is not offered merely as a parsimonious database for an external algorithm to scan. It is a model of self-activated memory. Each node is imagined as a small processor that maintains belief parameters for its proposition, monitors messages from neighbouring variables and revises when local compatibility is disturbed. A changed node can activate its neighbours, producing a wave of updates without a central controller scheduling every inference.

This architecture responds to cognitive constraints. Human short-term memory is limited; attention is narrow; many low-level interpretive tasks nevertheless feel fast and parallel. Pearl asks whether the knowledge representation itself can carry the pathways of computation. If links are the allowed communication channels and nodes repeatedly apply the same local operations, reasoning becomes distributed over the stored structure.

The proposal faces three problems that the paper states candidly. First, a stored link quantity such as P(A|B) constrains belief in A only in the special situation where B is certain and other evidence is absent. Current beliefs under a wider evidence set need not visibly satisfy that one number. Second, if evidence changes conditional beliefs, it appears that fixed link weights cannot support updating without continuous retuning. Third, predictive and diagnostic reinforcement seems capable of circular amplification: stronger belief in a cause predicts an effect, which then seems to strengthen the cause again without new evidence.

Pearl’s answer is not to send one undifferentiated probability around every edge. Each proposition carries a vector of support parameters, with components recording contributions from different directions or neighbours. Messages exclude the support that arrived from the recipient before returning an update. This decomposition prevents the same evidence from being counted again merely because it has travelled around a local exchange. Link weights remain fixed; the messages change as evidence changes.

The decomposition also has an explanatory benefit. A node can preserve a breakdown record of where its support came from. If a conclusion becomes surprising or wrong, the system can trace contributing assumptions and revise the model. This is not a complete modern explanation facility, but it anticipates a durable idea: inference should retain provenance rather than emit only a posterior number.

Provenance is especially important when reports are correlated. Two witnesses who copied the same bulletin do not supply two independent likelihood ratios. If their common source is absent from the graph, multiplying both messages exaggerates confidence while every local calculation appears correct. A source-aware network either represents the shared ancestor, models dependence directly, or consolidates the reports before propagation. The visual sparsity of a graph is not permission to assume independence wherever documentation is inconvenient.

The same distinction matters for conflicting evidence. A positive finding and a negative test do not “cancel” by verbal intuition. Each contributes a likelihood ratio defined by sensitivity and false-positive behaviour. Their product modifies prior odds. If one source is more diagnostic, the posterior moves toward it; if their reliabilities are uncertain, that uncertainty belongs in the model. The architecture replaces argument by adjectives with an inspectable composition of support.

“Autonomous” does not mean unconstrained. Every processor follows probability-preserving local equations. Self-activation determines when and where computation occurs; it does not license arbitrary belief revision. The equilibrium is globally coherent because the message rules implement the same joint distribution defined by the network: within the topology for which the algorithm is exact.

Singly connected networks make local propagation exact

The clean case is a singly connected network: between any two nodes there is at most one undirected path. It includes ordinary trees and polytrees with multiple-parent nodes so long as no pair is joined by two distinct routes. In that topology a message crosses each edge without another path carrying the same evidence back around. Local exclusion rules are enough to prevent duplication.

Pearl reports three properties. First, new information diffuses through the network in a single pass, and equilibrium time is proportional to the graph’s diameter rather than its total number of joint assignments. Second, primitive processors are simple and repetitive, requiring little working memory beyond their local calculations. Third, the final belief distribution is independent of the control mechanism: messages may be activated by data, goals, a clock or random scheduling and still reach the same equilibrium.

Schedule independence is a remarkable systems property. It permits asynchronous, event-driven implementation: a node can update when a neighbour changes rather than wait for a master traversal. Parallel hardware becomes natural because distant parts of the network can work simultaneously. The graph functions as both knowledge model and concurrency plan.

Diameter is the relevant latency measure in this idealized account because evidence must traverse the longest necessary route, while many branches can update in parallel. Total work still depends on node cardinalities and conditional-table sizes. A tree with one child having thousands of parent configurations may be topologically simple yet numerically expensive. Structural sparsity and small local state spaces are separate resources; production design must budget both.

Queries also affect activation. A data-driven system may push every observation outward immediately. A goal-driven system may request only messages needed for one target. The final exact posterior can agree while intermediate work differs greatly. Pearl’s schedule result therefore supports an engineering optimization: preserve semantic equations, then choose control policies that reduce unnecessary activation, respect latency priorities or fit parallel hardware.

The claim has a precise boundary. Schedule independence assumes valid local message equations, eventual delivery and a singly connected graph. It does not say that arbitrary neural activation, delayed stale messages or lossy numerical approximations are harmless. Nor does it establish that biological cognition implements these processors. Pearl offers the architecture as a viable formalism meeting connectionist computational objectives, not a completed neurological measurement.

Loops turn innocent local exchange into duplicated evidence

A directed acyclic graph can still contain undirected loops. Two variables may be connected by more than one route even though no arrow cycle exists. In such a multiply connected network, treating every node as if the graph were a tree lets the same evidence return along another path. Messages can circulate indefinitely, oscillate or settle at a value that is not the intended posterior. Local consistency is no longer sufficient for global correctness.

Pearl surveys four responses rather than pretending the problem disappears. Small loops can be collapsed into one compound node, restoring a tree at the cost of a larger state space. Stochastic relaxation can sample local states and converge in the long run, often slowly. Conditioning can instantiate a selected cutset variable, breaking loops for each case; exact tree propagation runs under each assumption and the results are recombined by the case probabilities. Auxiliary variables can restructure dependencies so that visible variables become leaves around hidden summarizing nodes.

Conditioning has a cognitive interpretation. Choose a key proposition, assume each of its possible values in turn, derive the consequences inside the simplified network, then integrate across cases. Pearl associates this with hypothetical reasoning, reasoning by cases and envisioning. The cost is exponential in the number and cardinality of conditioning variables, so the art lies in choosing a small cutset that breaks all relevant loops.

Each loop remedy preserves a different virtue. Clustering keeps exact deterministic propagation but can create tables whose size grows exponentially with cluster width. Sampling keeps local state representations but trades exactness for Monte Carlo error and convergence diagnostics. Conditioning keeps tree machinery but repeats it across cutset assignments. Auxiliary variables can reveal elegant structure, yet choosing them is a modelling act with interpretive consequences. There is no topology-free algorithmic lunch.

This menu foreshadows a broad principle of probabilistic computation: complexity is controlled less by the raw count of nodes than by how dependencies knit them together. A large sparse network can be easy; a modest dense core can dominate the calculation. Reporting only “number of variables” hides the operational risk. Report induced cluster size, cutset size, effective sample size or another method-appropriate measure of dependency complexity.

Auxiliary variables preserve uniform local processors but change the ontology of the model. A hidden variable can summarize an interaction among visible variables and render them conditionally independent. Pearl draws an analogy to the hypothetical constructs people invent: mental states, plans, disorders, particles or other unobserved causes: to organize observations into a causal hierarchy. This is a computational suggestion, not evidence that every useful latent variable names a real entity.

The 1985 paper sits at an important boundary. Its arcs are described as causal influences, and causal organization helps make diagnostic and predictive reasoning intelligible. Yet a modern system that learns a DAG from observational data cannot call every arrow causal by default. Statistical equivalence, hidden confounding and selection can leave direction underdetermined. The graph’s causal reading requires assumptions or interventions beyond factorization alone.

Nor should “self-activated memory” be inflated into a claim about consciousness. The paper proposes a computational model for aspects of human inferential reasoning and connectionist knowledge processing. It describes local activation, parallel communication and globally coherent belief revision. None of those properties, singly or together, establishes subjective experience. The relevant contribution to an intelligence-and-consciousness canon is architectural: it shows how structured memory can become active inference without becoming a central homunculus.

A protocol for active probabilistic memory

First, define propositions and state spaces before drawing edges. Ambiguous nodes make both conditional tables and evidence semantics unstable. Second, state whether an arrow is merely factorizing, temporally suggestive or genuinely causal. Third, write the local conditional table for every node and verify that the product normalizes. Fourth, list the conditional independencies implied by missing edges and challenge the ones that matter to action.

Fifth, classify topology. A polytree can use exact local propagation directly; a multiply connected graph needs an explicit method and complexity account. Sixth, define evidence handling, including hard observations, uncertain reports and missing values. Seventh, preserve message provenance so correlated sources are not multiplied as if independent. Eighth, test schedule invariance, normalization and equality to brute-force enumeration on small networks.

Ninth, make loop treatment visible: cluster size, cutset cardinality, sampling convergence or approximation error. Tenth, monitor the model boundary. Conditional tables age, interventions change data-generating processes, and hidden causes can become inadequate summaries. A self-activated network is not self-validating. Its local autonomy needs global audit.

Finally, design for correction. Keep versions of graph structure and parameters, record the evidence state behind each posterior, and distinguish recalculation from model revision. When a result is challenged, the team should be able to reproduce the messages, identify the path by which evidence arrived, remove one source, and show the counterfactual posterior. An active memory earns trust not by sounding certain but by making its own activation history inspectable.

Executable evidence-propagation laboratory

The laboratory uses a three-node tree C→F and C→T. Cause C has prior 0.2. Findings F and T are conditionally independent given C. Choose an evidence pattern. The engine computes the posterior by exhaustive enumeration and by multiplying local likelihood messages into prior odds, predicts any unobserved sibling, and checks normalization, equality and evidence-order invariance.

Memory studio