Hebbian learning

Hebbian learning is the oldest idea in computational neuroscience that is still in daily use: connections between units that are active together get stronger. It is usually summarised as "cells that fire together wire together", which is memorable, slightly inaccurate, and a good enough starting point.

The rule

Donald Hebb proposed in 1949 that when one cell repeatedly takes part in firing another, the connection between them strengthens. In computational form the simplest version is a product: the change in a weight is proportional to the activity at the sending end multiplied by the activity at the receiving end.

What makes this attractive is that it is entirely local. A synapse changes using only information available at that synapse. It needs no error signal, no correct answer, and no knowledge of what the rest of the network is doing, which is exactly the property biological plausibility demands and backpropagation lacks.

Eligibility traces and the timing problem

A second problem is timing. The consequence of a decision often arrives long after the decision. If a weight only changes when the two ends are active at the same instant, there is no way to connect an outcome now to the activity that caused it several steps ago.

The standard answer is an eligibility trace: each connection keeps a fading record of its own recent co-activity. The trace marks a connection as recently involved without yet changing it. When a reinforcement signal arrives, connections are adjusted in proportion to how eligible they still are. Recent contributors get most of the credit, older ones progressively less, and the whole thing stays local to the synapse.

Combining the two gives reward-modulated Hebbian plasticity: co-activity sets eligibility, and a reinforcement signal decides whether eligibility becomes a strengthening or a weakening. This is the family of rules with the best current claim to describing how dopamine-modulated learning works in real brains.

How it is used in Emergent Life

Organisms here have a within-life plastic component sitting alongside their inherited weights, driven by exactly this kind of rule: eligibility traces on each connection, gated by a reinforcement signal. The signal is not supplied by the simulation. It is computed by the organism from changes in its own internal state, and the parameters shaping it are themselves inherited and mutable, so how an individual learns is under the same selection as everything else about it.

Acquired weights start at zero at birth and are discarded at death. Nothing acquired is inherited, which is tested directly rather than assumed. The full architecture and the measured results, including two pre-registered null results reported alongside the one real effect, are on lifetime neural plasticity.