Synaptic scaling

Synaptic scaling is a form of homeostatic plasticity in which neurons uniformly scale the strength of all their synapses up or down to maintain a target level of overall activity. It was first characterized by Gina Turrigiano and colleagues in the late 1990s, who demonstrated that cortical neurons chronically deprived of activity increase synaptic strengths across all inputs, while neurons with excessive activity decrease them. This global scaling mechanism prevents runaway excitation or quiescence that would result from purely local Hebbian rules, acting as a slow-acting thermostat for neural circuits.

Synaptic scaling operates on timescales of hours to days — much slower than Hebbian long-term potentiation — and it is mediated by changes in postsynaptic receptor density rather than presynaptic release probability. The mechanism reveals that neural circuits are not merely collections of individually modifiable connections but regulated systems with global feedback loops. Without synaptic scaling, learning would be self-terminating: every Hebbian strengthening would push the network toward saturation, and every weakening toward silence. Scaling preserves the dynamic range.

The existence of synaptic scaling exposes a design principle that engineered neural networks have largely ignored: local learning rules require global homeostatic regulation. Deep learning systems that normalize activations across layers are converging on this principle independently, but they do so algorithmically rather than materially. A brain that scales its synapses is a brain that knows its own state. Current artificial systems do not.