Hebbian learning · Grey Matter

Hebbian learning is the rule that a connection between two neurons grows stronger when the first repeatedly helps to fire the second, and it is the simplest account of how experience writes associations into the wiring.

Hebbian learning. Hebbian learning is the rule that a connection between two neurons grows stronger when the first repeatedly helps to fire the second, and it is the simplest account of how experience writes associations into the wiring.

Donald Hebb proposed it in 1949, in The Organization of Behavior: when an axon of cell A is near enough to excite cell B and repeatedly takes part in firing it, some change occurs that increases A's efficiency in firing B. The popular summary, cells that fire together wire together, came later (from Löwel and Singer in 1992). The rule is local, since each synapse needs only the activity of its own two cells, and that is what makes it plausible for a brain with no central teacher. Its molecular form is the NMDA receptor and long-term potentiation.

Timing refines it. In spike-timing-dependent plasticity, a presynaptic spike arriving a few milliseconds before the postsynaptic spike strengthens the synapse, and one arriving after weakens it, within a window of about 20 ms; causality, more than mere coincidence, is what counts.

On its own it runs away. Strengthened synapses make the cells fire together more, which strengthens them further, so real circuits pair Hebbian plasticity with LTD, homeostatic scaling and inhibition that keep activity bounded.

A third factor gates it. A neuromodulator such as dopamine arriving after the coincidence can decide whether the change is kept, which links a local rule to a global signal of success.

Hebb's rule learns correlations, and a third factor makes it learn what matters.

Coincidence marks the candidate synapses; neuromodulators decide which of them become memory.

Questions: How can dopamine turn a Hebbian rule into learning from reward? A pure Hebbian rule strengthens every connection that takes part in firing a cell, whether the result was useful or not. In three-factor models, coincident activity of two cells only leaves a temporary mark on the synapse (an eligibility trace, lasting around a second), and the change becomes lasting only if a third signal arrives while the mark remains. Dopamine bursts after a better-than-expected outcome can play that role, and in the striatum, dopamine acting on D1 receptors shortly after a pre-post pairing has been shown to convert it into potentiation. The rule then strengthens the connections that preceded good outcomes, which is what reinforcement learning needs. Is long-term potentiation the same thing as Hebb's rule? Hebb's rule (1949) is a principle: a connection that repeatedly helps fire a cell is strengthened. LTP, found by Bliss and Lømo in 1973, is a measured phenomenon: synapses become stronger for hours after strong activity. NMDA-dependent LTP has the properties the rule asks for (it needs presynaptic release and postsynaptic depolarisation together, and it is specific to the active synapses), so it is the best candidate mechanism for Hebbian learning. They are still distinct: some forms of LTP are not Hebbian, and the rule as stated says nothing about weakening, which LTD and spike-timing rules add. Why does the order of two spikes, a few milliseconds apart, decide whether a synapse strengthens or weakens? In spike-timing-dependent plasticity, a presynaptic spike arriving a few milliseconds before the postsynaptic spike strengthens the synapse, and one arriving just after weakens it, within a window of about 20 ms on each side in Bi and Poo's hippocampal cultures. The NMDA receptor explains much of it: if glutamate binds first and the postsynaptic spike then back-propagates into the dendrite and expels the magnesium, a large burst of calcium enters, which favours potentiation. If the postsynaptic spike comes first, the depolarisation has passed by the time glutamate arrives, calcium entry is small, and depression follows. The rule rewards inputs that could have caused the spike and punishes those that came too late, a causal refinement of Hebb's idea.