In traditional artificial neural networks, a synaptic weight is a static number—a scalar that changes only during a slow "learning" phase. In biological brains, however, synapses are dynamic filters. Their efficacy fluctuates on millisecond timescales based on the history of activity. This phenomenon, known as Short-Term Plasticity (STP), is not a bug or a sign of "fatigue," but a fundamental computational mechanism that allows neural circuits to process time, maintain stability, and generate complex rhythms.
The Synapse as a Finite Resource
The core intuition behind short-term plasticity—specifically Short-Term Depression (STD)—is resource depletion. A synapse relies on vesicles filled with neurotransmitters. When a neuron fires, some of these vesicles are released. If the neuron fires again quickly, there are fewer vesicles available, and the resulting signal is weaker. Over time, the "pool" of vesicles replenishes.
Thomas Natschläger, Wolfgang Maass, and colleagues modeled this as a time-varying product of a base weight, a facilitation factor (which increases with activity), and a depression factor (which decreases as resources are used) in Processing of Time Series by Neural Circuits with Biologically Realistic Synaptic Dynamics. This simple kinetic model transforms a synapse from a static multiplier into a frequency-dependent filter.
Self-Organized Stability and Criticality
One of the most profound roles of STP is keeping the brain in a "Goldilocks zone" of activity. If synapses were static, a small burst of activity could easily spiral into runaway excitation (like an epileptic seizure) or die out into total silence.
Anna Levina and J. Michael Herrmann demonstrated in Dynamical synapses causing self-organized criticality in neural networks that STD acts as a built-in negative feedback loop. When the network becomes too active, synapses weaken, naturally dampening the fire. When activity drops, synapses recover, making the network more sensitive. This allows the system to self-organize into a state of criticality, characterized by "avalanches" of activity that follow a power-law distribution—a state that information theory suggests is optimal for information transmission and processing.
Moving Attractors and Delay Compensation
In models of spatial navigation or visual tracking, activity is often represented as a "bump" in a Continuous Attractor Neural Network (CANN). With static synapses, this bump stays where it is unless pushed by an external input. However, Si Wu and C. C. Alan Fung showed that adding synaptic depression changes the game.
In Attractor Dynamics with Synaptic Depression, they explain that as neurons at the center of an activity bump fire, their synapses fatigue. This makes the "trailing edge" of the bump weaker than the "leading edge," effectively pushing the bump forward. This mechanism allows the brain to track moving stimuli more efficiently and even perform delay compensation, anticipating where a fast-moving object will be despite the slow processing time of biological neurons (Delay Compensation with Dynamical Synapses).
Generating Rhythms through Mutual Inhibition
Locomotion and breathing require rhythmic, alternating signals. A classic model for this is the "half-center oscillator," where two neurons (or populations) inhibit each other. But how does the system switch? If Neuron A is inhibiting Neuron B, how does Neuron B ever get a chance to fire?
Synaptic depression provides the "off-switch." As Neuron A fires, its inhibitory synapses onto Neuron B gradually weaken due to resource depletion. Eventually, the inhibition becomes so weak that Neuron B is released from suppression and begins to fire, which in turn inhibits Neuron A. Conor Houghton and Mark Olenik explored this in A scalar Poincaré map for struggling in Xenopus tadpoles, showing that this temporal weakening is a primary driver for the stable, anti-phase bursting required for central pattern generators.
Protecting Memory and Detecting Novelty
STP also plays a role in higher-level cognition. Tomoki Fukai argued in Interplay between Short- and Long-Term Plasticity in Cell-Assembly Formation that short-term depression prevents neural assemblies from staying active indefinitely. By forcing these assemblies to cycle between active and silent states, STP prevents runaway excitation from washing away long-term memory traces.
Furthermore, Xiuning Zhang and Yuanyuan Mi have shown that depression in early sensory layers acts as a novelty detector. By weakening the response to repeated, familiar stimuli, the circuit creates a "repulsive bias" that highlights new or changing information in a dynamic scene (Neural Correlates of Serial Dependence: Synaptic Short-term Plasticity Orchestrates Repulsion and Attraction).
Summary of Computational Roles
| Mechanism | Computational Function | Key Researchers |
|---|---|---|
| Resource Depletion | Frequency-dependent filtering | Maass, Natschläger |
| Negative Feedback | Self-organized criticality & stability | Levina |
| Asymmetric Fatigue | Delay compensation & stimulus tracking | Wu, Fung |
| Inhibitory Weakening | Rhythmic pattern generation (CPGs) | Houghton |
| Activity Cycling | Memory protection & novelty detection | Fukai, Zhang |
By treating synapses as dynamic entities rather than static weights, these researchers have shown that the "limitations" of biological hardware—like depletion and fatigue—are actually sophisticated tools for temporal computation.