Why correlation alone does not create intelligence—and why most explanations stop too early

Keywords: Hebbian learning, reinforcement learning, prediction error, dopamine learning, neural networks, emergence in AI, three-factor learning rule, temporal difference learning, AI theory, neuroscience AI

Maurício Pinheiro

There is something deceptively elegant—almost suspiciously elegant—about the phrase “neurons that fire together wire together.” It has the rhetorical efficiency of a slogan and the intellectual danger of one. Repeated often enough, it begins to masquerade as a complete explanation, when in reality it is only the opening move of a much longer argument.

And that is precisely where most explanations fail.

Hebbian learning does not explain intelligence. It explains how structure emerges from coincidence. Confusing the two is not a minor conceptual mistake—it is the reason why so many narratives about artificial intelligence sound convincing while remaining fundamentally incomplete.

To understand what is actually happening, we must move beyond the slogan and into the mechanism.

At its core, Hebbian learning is a local synaptic update rule, first proposed by Donald O. Hebb in his 1949 book The Organization of Behavior. The formal expression is deceptively simple:$$\Delta w_{ij} = \eta \, x_i \, y_j$$
This equation encodes an entire philosophy of learning. The change in synaptic weight $\Delta w_{ij}$depends only on three quantities: the activity of the pre-synaptic neuron $x_i$xi​, the activity of the post-synaptic neuron $y_j$​, and a learning rate $\eta$. There is no global objective, no supervision, no notion of correctness—only co-activation.

This is not optimization. It is correlation detection.

The system strengthens connections when signals co-occur. Over time, it becomes a statistical memory of the environment—a map of what tends to happen together. In modern terms, Hebbian learning approximates the extraction of second-order statistics, closely related to covariance structures and unsupervised feature learning.

But here lies the first fracture in the story: correlation is not meaning.

The world is saturated with correlations. Many are useful, many are irrelevant, and some are actively misleading. Hebbian learning does not distinguish between them. It cannot. There is no term in the equation that encodes importance, utility, or truth. It reinforces repetition—blindly.

This is why pure Hebbian systems converge toward representation, but not toward intelligence. They accumulate structure, but they lack direction.

Historically, this limitation became evident as early as the 1960s, when researchers began exploring associative memory models. While Hebbian-like rules could store patterns (as later formalized in John Hopfield networks in 1982), they could not decide which patterns should be stored. The system had memory, but no criterion for relevance.

What Hebbian learning lacks is value.

And this is where reinforcement enters—not as an extension, but as a transformation.

When a scalar evaluation signal is introduced, the learning rule becomes:

$$\Delta w_{ij} = \eta \, x_i \, y_j \, R$$

This is the essence of the three-factor learning rule, widely studied in both neuroscience and machine learning. The additional term $R$R represents a reinforcement signal—reward, punishment, or evaluative feedback.

With this single modification, the system acquires something profoundly new: selectivity.

Now, correlation is no longer sufficient for learning. It must also be validated by outcome. Connections are strengthened only when co-activation coincides with positive reinforcement, and weakened when associated with negative outcomes.

In biological systems, this mechanism is deeply tied to dopaminergic signaling. The seminal work of Wolfram Schultz in the 1990s demonstrated that dopamine neurons encode reward-related signals, effectively acting as a global reinforcement broadcast. This transforms local Hebbian plasticity into a reward-modulated learning system.

In artificial systems, the same principle underlies reinforcement learning, formalized by Richard S. Sutton and Andrew G. Barto. Here, agents learn policies not by memorizing correlations, but by reinforcing action-state pairs that maximize expected reward.

The difference is subtle, but decisive: Hebbian learning answers “what occurs together?” Reinforcement learning answers “what leads to success?”

But even this is not enough.

Because reinforcement alone treats all rewards equally—it lacks temporal nuance. Biological and artificial systems go further by incorporating prediction error, a concept that bridges neuroscience and machine learning.

Formally:$$\delta = r – \hat{r}$$

Where $r$ is the received reward and $\hat{r}$ is the expected reward. This quantity, known as the reward prediction error, is the true driver of learning.

This idea is central to Temporal Difference (TD) learning, one of the cornerstones of modern reinforcement learning. It is also, remarkably, mirrored in the brain. Schultz’s experiments showed that dopamine signals do not encode reward itself, but deviations from expectation.

Unexpected reward → strong positive update.
Expected reward → minimal update.
Missing expected reward → negative update.

Learning, therefore, is not about reinforcement per se—it is about surprise.

This transforms the learning process into a continuous loop: prediction, action, outcome, correction. The system no longer passively records patterns; it actively refines its expectations about the world.

At this stage, something qualitatively new emerges.

The network begins to encode not just correlations or rewards, but models. These models are implicit—distributed across weights—but they capture regularities about both the environment and the consequences of action.

This is the true architecture of emergence: a layered interaction between correlation, valuation, and error correction.

Each component is necessary, and none is sufficient on its own. Remove correlation, and there is no structure. Remove reinforcement, and there is no direction. Remove prediction error, and there is no adaptation.

Together, they form a minimal substrate from which intelligence can arise—not as a predefined property, but as an emergent consequence.

This perspective also reframes modern deep learning. Backpropagation, often presented as a purely mathematical optimization procedure, can be interpreted as a global approximation of these same principles. It propagates error signals backward through the network, aligning local updates with global objectives. Yet even here, the tension remains: biological systems rely on local rules, while artificial systems exploit global gradients.

The bridge between these paradigms—how local plasticity approximates global optimization—remains one of the deepest open questions in both neuroscience and AI.

And perhaps the most uncomfortable implication lies at the end of this chain.

If learning is driven by reinforcement, then what persists is not necessarily what is true—but what has been consistently rewarded.

Neural systems—biological or artificial—do not converge to truth. They converge to reinforced regularities.

This applies as much to a child learning to catch a ball as it does to an algorithm optimizing engagement metrics. Patterns stabilize because they are rewarded, not because they are inherently correct.

Which leads to a final compression of the entire process:

Learning is correlation, filtered by value, corrected by error, and stabilized through repetition.

Everything else—perception, reasoning, abstraction—is built on top of this.

And from this perspective, intelligence becomes less of a mystery and more of a process. Not something inserted into a system, but something that emerges when structure, value, and correction begin to interact.

We do not simply learn what is true.

We learn what survives reinforcement.

---

📚 References

• Hebb, D. O. (1949). The Organization of Behavior — https://archive.org/details/in.ernet.dli.2015.177405
• Hebbian Theory — https://en.wikipedia.org/wiki/Hebbian_theory
• Hopfield, J. (1982). Neural networks and physical systems — https://www.pnas.org/doi/10.1073/pnas.79.8.2554
• Schultz, W. (1997). Dopamine and reward — https://www.science.org/doi/10.1126/science.275.5306.1593
• Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning — http://incompleteideas.net/book/the-book.html
• Three-Factor Learning Rule — https://en.wikipedia.org/wiki/Three-factor_learning
• Lillicrap et al. (2020). Backpropagation and the brain — https://www.nature.com/articles/s41583-020-0277-3

#ArtificialIntelligence #MachineLearning #Neuroscience #ReinforcementLearning #DeepLearning #Emergence #AITheory #HebbianLearning #AIResearch #AITalks

---

Copyright 2026 AI-Talks.org