Before Donald Hebb, a neuron was a fixed thing. The McCulloch–Pitts model (1943) had shown that neurons could compute Boolean logic, but their connections were frozen at construction. A network could think, but it could not learn.
Hebb changed that with a single principle. In The Organization of Behavior (1949), he wrote:
“When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A’s efficiency, as one of the cells firing B, is increased.”
Stripped to its core: neurons that fire together wire together.
The mechanism is local and simple. No global supervisor. No error signal. Just two neurons and a rule: if A helps fire B, strengthen the connection from A to B.
In mathematical form, stripped to its simplest, Hebbian learning says:
The weight from neuron to neuron increases in proportion to the product of their activations. If both fire (, ), the connection grows. If either is silent, nothing changes. The learning rate sets the step size.
This simple rule is unstable. Stronger connections drive more firing, which strengthens them further — a positive feedback loop with no bound. There is no competition between inputs, no mechanism for weight reduction, no forgetting. Left to itself, Hebbian learning runs away.
Later modifications — weight normalisation, Sanger’s rule, Oja’s rule — added stability. But the core insight survived: learning is local, connection-specific, and driven by correlation.
Hebbian learning is the bridge between the McCulloch–Pitts neuron and the connectionist machine. The McCulloch–Pitts neuron could compute but not change. Hebb gave it plasticity. The fixed logic gate became a modifiable weight, and the network became something that could learn from experience.