The Synapse

A chemical gap where a weight is stored, modified, and read — the biological parameter.

The structure

A synapse is a narrow cleft, roughly 20 nanometers, between a presynaptic terminal and a postsynaptic membrane. The presynaptic side holds vesicles of neurotransmitter and the machinery to release them on calcium influx; the postsynaptic side holds receptors, and a dense protein scaffold that anchors them. Excitatory synapses in cortex typically sit on dendritic spines — small protrusions whose neck geometry electrically and chemically isolates the synapse from the parent dendrite. Spine shape changes on a timescale of minutes.

The function it dictates

The synapse sets how strongly one neuron influences another, and that strength is the storage medium of learning. Strength depends on release probability, vesicle count, and receptor density — several partly independent variables rather than one number. Transmission is stochastic: a given spike may release nothing at all. Strength changes with correlated activity through long-term potentiation and depression, and it is gated by neuromodulatory state, so the same activity pattern produces different learning depending on context.

The mirror in AI

A synapse is a weight. The parallel is genuinely tight — this is the one place where the biological metaphor earns its keep. Hebbian learning ("fire together, wire together") maps onto correlation-based update rules, and dropout is a rough analogue of stochastic transmission, though it was invented for regularization rather than in imitation of anything.

The mirror in robotics

The closest control-theoretic analogue is an adaptive gain: a coupling coefficient tuned online by experience rather than fixed at design time. Adaptive control systems that adjust their gains from tracking error are doing, in a much simpler form, what synaptic plasticity does — and they face the same stability problem, since a system that modifies its own gains can destabilize itself.

Where the analogy breaks

An artificial weight is a single real number, updated globally, deterministic, and unbounded in either direction. A synapse is multi-variable, updated locally, stochastic, sign-constrained (a given synapse is excitatory or inhibitory and does not flip), metabolically expensive to maintain, and physically remodeled — spines appear and disappear, so the connectivity graph itself changes rather than only its weights. Artificial networks have a fixed graph and variable weights; brains vary both.

Open questions

Does synaptic stochasticity serve a computational purpose or is it a cost of chemistry?
How does the brain solve the credit assignment problem with only local signals?

In Kiru Lab

Values, Names, and Objects The Artificial Neuron and Its Honest Limits

Related DEM-X entries

GI-DRFT-01