Information Processing with Short Term Synaptic Depression

67 concepts47 papers2 proposals

In traditional artificial neural networks, connection strengths (weights) are static during inference. They change only during training via optimization algorithms. In the biological brain, however, a synapse is not a fixed scalar multiplier. It is a dynamic, history-dependent filter. When a presynaptic neuron fires rapidly, the strength of its connection to the postsynaptic neuron temporarily drops. This phenomenon is known as short-term synaptic depression (STD).

Rather than being a mere biological flaw or a sign of "neural fatigue," short-term depression serves as a fundamental computational mechanism. By allowing synapses to adapt on millisecond-to-second timescales, the brain solves complex problems in stability, temporal filtering, spatial tracking, and network coordination.

The Biophysical Intuition: Depletion and Recovery

Why does a synapse tire out? At a physical level, chemical neurotransmission relies on a finite pool of neurotransmitter vesicles ready to be released into the synaptic cleft. As modeled by Morrison et al., each presynaptic spike triggers the release of a fraction of these available resources.

If the neuron fires again almost immediately, the vesicle pool has not yet had time to replenish. Consequently, fewer vesicles are released, and the resulting postsynaptic impact is weakened (Moezzi et al.; He et al.). This resource exhaustion is often governed by the initial probability of neurotransmitter release: connections with a high initial release probability exhaust their vesicle pools rapidly, leading to pronounced short-term depression (McFarlan et al.). This interplay between activity-dependent depletion and passive temporal recovery forms the mathematical backbone of phenomenological frameworks like the Tsodyks-Markram model (Natschläger et al.; Prince et al.).

Decentralized Stability and Criticality

What happens when an entire recurrent network is constructed from these fatiguing connections? The network gains a powerful, decentralized negative feedback loop.

If a cluster of neurons becomes hyperactive, their outgoing synapses rapidly depress, automatically dampening the propagation of excitation. Conversely, when the network falls quiet, synapses recover their full resource pools, increasing their sensitivity to incoming signals. Levina et al. demonstrated that this activity-dependent feedback allows neural networks to autonomously self-organize toward a state of criticality—the optimal boundary between total silence and chaotic, runaway excitation—without requiring an external parameter tuner (see also Levina et al.; Levina et al.).

This automatic gain control stabilizes local circuits (Blackman et al.; Zhang et al.) and protects long-term memory traces. As Hiratani et al. showed, by forcing neural assemblies to cycle between active and silent states, short-term depression prevents spontaneous neural noise from eroding long-term synaptic weights over time.

Synapses as Temporal Filters

Because the degree of synaptic weakening depends strictly on the recent history of spikes, a depressing synapse acts as a frequency-dependent filter. As Senn showed, the probability of signal transmission drops as the presynaptic firing rate increases. High-frequency bursts cause rapid exhaustion, meaning the synapse responds robustly to the onset of a signal but attenuates sustained, redundant inputs.

This history dependence turns synapses into temporal "locks" that are uniquely tuned to specific spike patterns. Natschläger et al. demonstrated that the spike trains maximizing a synapse's response closely match the natural bursting patterns of real neurons, allowing networks to decode real-time time-series data like speech or motion (Wu et al.). From an information-theoretic perspective, this suppression of redundant high-frequency inputs maximizes information efficiency by subtracting out predictable internal fluctuations (Toyoizumi; Stevenson et al.).

Spatial Asymmetry and Motion Tracking

One of the most elegant applications of short-term depression occurs in Continuous Attractor Neural Networks (CANNs), which are widely used to model how the brain tracks continuous variables like a stimulus's position or orientation.

When a visual stimulus moves across a scene, it activates a sequence of corresponding neurons in the network. As the neurons at the center of the current activity peak fire intensely, their synapses begin to deplete and weaken. This creates a localized "fatigue trail" directly behind the moving bump of neural activity. Fung et al. demonstrated that this spatial asymmetry effectively pushes the center of neural activity forward toward less-depleted neurons. This mechanism allows the network to compensate for inherent biological processing delays and even anticipate future stimulus positions (see also Fung et al.; Wang et al.).

Dynamic Circuit Switching

Finally, short-term depression allows neural circuits to dynamically shift their global processing strategies over time without requiring physical rewiring.

In the primary visual cortex, Adorján et al. showed that fast synaptic depression causes the network to start in a highly competitive, nonlinear "winner-take-all" mode to rapidly lock onto salient features, and then automatically relax into a more linear, cooperative mode to process finer details as the dominant units fatigue.

A similar principle governs motor and rhythmic circuits. The temporal weakening of inhibitory connections provides a natural mechanism for driving alternating, out-of-phase bursting patterns required for locomotion and central pattern generation (Houghton et al.; Köhler et al.). By periodically reducing the suppression exerted by an active neuron on its partner, the circuit can seamlessly transition between active states purely as a function of time.

Key Foundational Work

Recurrent Cortical Competition: Strengthen or Weaken? (1999) — contributed 1 concept elements to this direction.
Stationarity of Synaptic Coupling Strength Between Neurons with Nonstationary Discharge Properties (1991) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 48 papers
  2. Version 1 · 2026-10-04 · 34 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_short_term_synaptic_depression_35c34f7a,
  title = {short-term synaptic depression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/short-term-synaptic-depression/rdr_35c34f7a429c4e449f7e489f953acceb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_short_term_synaptic_depression_35c34f7a_v2,
  title = {short-term synaptic depression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/short-term-synaptic-depression/rdr_35c34f7a429c4e449f7e489f953acceb/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake