The Organization of Behavior
Authors: Donald O. Hebb
Year: 1949 | Wiley
Source: hebb-1949-the-organization-of-behavior-a
What it is
Hebb’s 1949 book proposed the biological mechanism for learning and memory in the brain: when two neurons fire together, the connection between them strengthens. This “Hebbian learning rule” became the foundation of unsupervised learning in neural networks and, decades later, was confirmed experimentally as Long-Term Potentiation (LTP).
The postulate (Chapter 4)
From the book:
“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.”
Summarized: Cells that fire together, wire together.
In modern notation:
Δw_{AB} = η · x_A · x_B
- If A and B are simultaneously active (x_A · x_B > 0): strengthen the connection
- If one is inactive: no change (or weakening, in extended versions)
- No error signal, no teacher — purely local, unsupervised
The cell assembly
Repeated co-activation through Hebbian learning binds neurons into a cell assembly — a group that becomes strongly interconnected and tends to activate as a unit.
Properties of a cell assembly:
- Activation of any part tends to activate the whole (associative recall)
- Can sustain reverberant activity briefly after the stimulus ends (short-term memory)
- Permanent structural changes from repeated reverberation (long-term memory)
- Represents a concept, percept, or memory — “apple,” “red,” “mother”
Assemblies activate each other in learned sequences (phase sequences) — Hebb’s model of sequential thought and planning.
Why it was speculative in 1949
Hebb acknowledged that direct evidence for synaptic strengthening through co-activation was not available in 1949. Electron microscopy capable of visualizing synaptic knob changes didn’t yet exist. He was extrapolating from:
- Lorente de Nó’s anatomical work on reverberatory circuits
- Behavioral studies on perceptual learning
- Physiological reasoning about what would be necessary for memory to work
The postulate was confirmed experimentally 24 years later when Bliss and Lømo (1973) discovered Long-Term Potentiation (LTP) in the hippocampus: high-frequency stimulation of a synapse produces long-lasting strengthening — exactly what Hebb predicted.
Influence on machine learning
| Neural network concept | Hebb connection |
|---|---|
| Perceptron learning rule | Supervised Hebbian: reinforce when prediction is correct |
| Hopfield network (1982) | Content-addressable memory using Hebbian weights |
| Hebbian unsupervised learning | Principal component analysis via online Hebbian rule |
| Spike-timing dependent plasticity | Refined Hebb: timing matters (pre before post = LTP; post before pre = LTD) |
| Self-organizing maps | Competitive Hebbian learning |
| Contrastive Hebbian Learning | Boltzmann machines, Predictive Coding |
Modern deep learning does not directly use the Hebbian rule — backpropagation is global and requires an error signal. But Hebb’s insight that local co-activity can drive learning continues to inspire biologically plausible alternatives to backpropagation.
Hebbian learning is unstable
Pure Hebbian learning has a problem: weights grow without bound. If A and B fire together and strengthen their connection, they’ll fire together more, strengthening further… → runaway excitation.
Solutions:
- Weight normalization: cap maximum weight or normalize periodically
- Oja’s rule (1982): online Hebbian with weight decay — converges to the first principal component
- BCM theory (Bienenstock, Cooper, Munro 1982): sliding threshold separates LTP from LTD
- Synaptic homeostasis: the brain globally scales synaptic weights down during sleep
See also
- mcculloch-1943-a-logical-calculus-of-the — The neuron model Hebb built on
- rosenblatt-1958-the-perceptron-a-probabilistic-model — First learning algorithm (supervised Hebb)
- schultz-1997-a-neural-substrate-of-prediction — Dopamine: reward-modulated Hebbian learning