When Neurons Learn "Because": The Causal Awakening of Neural Assemblies
> Paper: Causal Learning with Neural Assemblies > Authors: Evangelia Kopadi, Dimitris Kalles > arXiv: 2604.26919 > Published: 2026-04-29 > Fields: Machine Learning / Neuroscience / Causal Reasoning
1. An Ancient Metaphor About Fire
Imagine you are a primitive human seeing fire for the first time. You notice regularities: lightning strikes a tree, the tree burns; the burning tree ignites nearby grass; burning grass makes small animals flee.
You have learned three things: 1. Lightning → tree burns 2. Tree burns → grass burns 3. Grass burns → animals flee
But what do you actually *know*? "The tree burns after lightning" is temporal precedence. "If the tree burns, the grass burns" is conditional dependence. Do you truly know that lightning causes the tree to burn? From the data alone, the reverse direction looks equally consistent.
This is the core problem of causal reasoning: correlation does not imply causation. The rooster crows before sunrise, but doesn't wake the sun. Ice cream sales correlate with drownings only because summer causes both.
Humans perceive causation almost instinctively—we see someone push a door and immediately perceive the push as the cause. But what about AI?
2. Neural Assemblies: The Brain's "Choir"
A key concept underpinning the paper is the neural assembly. In conventional neural networks, information is carried by individual neuron activations—like individual lamps switching on and off. In real brains, information often appears as groups of neurons activating together, like a choir section singing in unison. When you think of "apple," no single "apple neuron" fires; a distributed group fires—color (red), shape (round), taste (sweet), emotion (liking).
Donald Hebb proposed in 1949 the famous principle "neurons that fire together, wire together." If a stimulus repeatedly activates the same group of neurons, their mutual synapses strengthen until they function as a single unit—an assembly.
Papadimitriou et al. (2020) proved a milestone result: neural assemblies are computationally universal—given enough of them, they can implement any Turing-computable function. This gives a more biologically grounded foundation than typical mathematical abstractions.
But one question remained open:
**Can neural assemblies learn causal *direction*?
If one assembly represents "lightning" and another "tree burning," can they learn "lightning causes burning" rather than the reverse? This matters because assemblies promise biological plausibility—they rely on local, Hebbian plasticity rather than global backpropagated gradients. If they can solve causal learning, we might build causal AI that is biologically grounded and mechanically interpretable.
3. Why Is Direction So Hard to Learn?
The difficulty is underestimated. Most causal methods require interventions (do-calculus) or strong assumptions (e.g., no hidden variables). Assemblies face stricter constraints: they can only learn from observation and local plasticity—no active experiments, no global information.
Suppose assemblies A and B frequently co-activate. Possible explanations: 1. A causes B (A → B) 2. B causes A (B → A) 3. Hidden C causes both (C → A, C → B)
From co-occurrence data alone, these are indistinguishable—the fundamental difficulty formalized by Judea Pearl in the 1990s. Pearl's remedy is the do-operator: intervene on a variable and observe. But assemblies cannot run experiments; they learn passively. It's like someone who has read all their life that "smoking" and "lung cancer" co-occur but can never run a randomized trial to disentangle the direction or confounders.
4. DIRECT: Giving Assemblies a Sense of Direction
The paper's core contribution is DIRECT (DIRectional Edge Coupling/Training).
Core intuition: activation order and strength are asymmetric. If two things are merely correlated, their connection should be roughly symmetric. But if A causes B:
1. Temporal asymmetry: A activates first, B follows. 2. Strength asymmetry: the A→B synapse should be stronger than B→A, like a one-way street. 3. Propagation asymmetry: signal flows smoothly from A to B but is impeded from B to A.
DIRECT achieves this via "co-activation + adaptive gain scheduling": the "cause" and "effect" assemblies are co-activated during training, but their connecting synapses receive an asymmetric gain. Like a child internalizing the difference between pushing and pulling through repeated embodied experience, the assemblies physically encode direction as synaptic-weight asymmetry.
Three key operations—all native to the assembly framework, with no biologically implausible additions:
1. Projection: information passes from one assembly to another. 2. Local Plasticity Control: each synapse adjusts only from its own and neighboring activity—no global supervision, like musicians coordinating without a conductor. 3. Sparse Winner Selection: only the strongest signals propagate onward, like a tournament where only the champion advances—preserving signal purity against noise.
The keyword is local. No global gradient flows backward; each synapse adjusts from local activity—the essence of Hebbian learning and, plausibly, of real brains.
5. How Do We Know It Actually Learned?
The paper uses an elegant dual verification strategy:
Verification 1: Synaptic-Strength Asymmetry. If the assembly learned "A causes B," then A→B weights should significantly exceed B→A. The measured "weight gap" confirms this after DIRECT training. Crucially, this checks the mechanism, not the performance—not "how many answers were right" but whether the internal physical structure embodies directionality.
Verification 2: Functional Propagation Overlap. Beyond static weights, the paper examines dynamic activation flow. Signal injected at A should reliably reach B; signal injected at B should not reliably flow back to A. Forward propagation reliability significantly exceeds reverse.
This recalls Feynman's Brazilian students who could recite every electromagnetism formula but couldn't answer simple physical questions—memorizing is not understanding. The dual verification guards against such cargo-cult learning by inspecting internal mechanism rather than test scores.
Results: In the supervised, known-structure setting (where "A→B" is given as correct during training), DIRECT achieves perfect structural recovery—precisely reconstructing the hidden causal graph. But "perfect" comes with the caveat of an idealized setting; the paper does not report performance under unknown structure or noise.
6. Feynman's Interrogation
Is "local plasticity" a sleight of hand?** If plasticity is local but the *timing and magnitude* of the gain schedule are coordinated by a global signal, that is implicit global supervision—"local" becomes marketing language. The paper does not fully clarify whether gain scheduling is entirely local.
Is perfect structural recovery overfitting? Answering correctly when the answer is known proves little. The paper reports no results under unknown structure or noisy data—conditions that define the real world. A system perfect in the lab may fail in the wild.
Assemblies vs. conventional networks. Is the claimed "explainable by design" genuine interpretability? A large assembly's collective behavior may still be complex and emergent. "Auditable" is not the same as "understandable." Still, synaptic asymmetry and propagation overlap as metrics crack open the black box further than conventional networks allow.
Cargo-cult detection. "Because it resembles the brain, it must be good" is itself cargo-cult reasoning. The brain is an evolved bundle of compromises. DIRECT's value should rest on what it can do and why it works—which the paper does address experimentally—but the word "assemblies" carries rhetorical force that can lull critical thinking.
7. Toward Interpretable Causal AI
Current AI systems—especially large language models—are correlation machines, not causal machines. This flaw is dangerous in:
- Medical diagnosis: recommending treatment for symptom A when A merely indicates hidden root cause C can lead to misdiagnosis.
- Policy making: "areas with more police have more crime" might suggest cutting police, when causality runs the other way.
- Science: models can extract "A correlates with B" from literature but cannot generate "A causes B" hypotheses for experiments.
A deeper philosophical question: what would the assembly actually have learned—a statistical regularity ("A precedes B, so A→B") or genuine mechanistic understanding ("A activates a mechanism that produces B")? From the experimental design, DIRECT appears closer to the former. True causal understanding may require knowing *why* A causes B, not just *that* it does.
Feynman might say: "This is like a child who has just learned to walk. You can't say they'll never run a marathon—but you can't call them a marathon runner yet either."
References
1. Kopadi, E., & Kalles, D. (2026). *Causal Learning with Neural Assemblies*. arXiv:2604.26919. 2. Hebb, D. O. (1949). *The Organization of Behavior*. Wiley. 3. Papadimitriou, C. H., et al. (2020). *Brain Computation by Assemblies of Neurons*. PNAS, 117(25), 14464-14472. 4. Pearl, J. (2009). *Causality: Models, Reasoning, and Inference* (2nd ed.). Cambridge University Press. 5. Feynman, R. P. (1974). *Cargo Cult Science*. Caltech Commencement Address. 6. Kahneman, D. (2011). *Thinking, Fast and Slow*. Farrar, Straus and Giroux.
*Analysis completed 2026-05-01 | Feynman-perspective calibration passed*