Homeostasis and Learning through Spike-Timing Dependent Plasticity

L. F. Abbott, Wulfram GerstnerInfoscience 2003

Abstract

Synaptic plasticity is thought to be the neuronal correlate of learning. Moreover, modification of synapses contributes to the activity-dependent homeostatic maintenance of neurons and neural networks. In this chapter, we review theories of synaptic plasticity and show that both homeostatic control of activity and detection of correlations in the presynaptic input can arise from spike-timing dependent plasticity (STDP). Relations to classical rate-based Hebbian learning are discussed.

Introduction: The Stability-Plasticity Dilemma

In the quest to understand how the brain learns, researchers have long grappled with the "stability-plasticity dilemma." Traditional Hebbian learning, often summarized by the phrase "cells that fire together, wire together," presents a significant challenge: it is inherently unstable. When a synapse strengthens due to high activity, it further increases the likelihood of future activity, creating a runaway positive feedback loop. Without intervention, this process would drive neurons to their maximum firing rates and saturate all available synapses, rendering the network non-functional.

Historically, theorists addressed this by imposing global constraints, such as manually forcing the sum of all synaptic weights to remain constant. However, this paper explores whether a local rule—acting independently at each synapse—can provide both the sensitivity required for learning and the stability necessary for homeostasis without requiring global oversight.

STDP: A Dual Mechanism for Neural Circuits

Spike-Timing Dependent Plasticity (STDP) emerges as a powerful "two-in-one" mechanism. Unlike classical rate-based models that rely on average firing frequencies, STDP operates on the millisecond-level timing between individual pre- and post-synaptic spikes.

Schematic of synaptic transmission (A) and the arrival of an action potential (AP) triggering neurotransmitter release (B). STDP rules depend on the precise timing of these spikes.

By focusing on the precise timing of events, STDP allows neurons to detect causal correlations in their inputs. If a presynaptic spike consistently precedes a postsynaptic action potential, the synapse is strengthened (Long-Term Potentiation, LTP). Conversely, if the presynaptic spike follows the postsynaptic spike, the synapse is weakened (Long-Term Depression, LTD). This sensitivity to order allows the circuit to learn causal relationships rather than just simple coincidences.

Theoretical Framework and Mean Field Approach

To bridge the gap between spike-based and rate-based models, the authors utilize a Mean Field Approach. By integrating the STDP "window function"—the curve defining synaptic change relative to spike timing—against the correlation functions of spike trains, they derive a weight evolution equation. This mathematical derivation demonstrates that STDP is not a departure from classical theory but rather a refinement of it, relating directly to established rules like Oja’s rule and the Bienenstock-Cooper-Munroe (BCM) rule. Under this approximation, the complex spike-timing rule reduces to a version of Sejnowski’s covariance rule, providing a unified theory for learning across scales.

Key Insights: Homeostasis and Normalization

One of the most significant findings is that STDP naturally implements intrinsic subtractive normalization. Because the typical STDP window allocates a larger area for depression than for potentiation, synapses that do not contribute to causal firing are systematically weakened. This introduces competition between synapses without the need for a global "summing" rule.

Furthermore, the authors demonstrate the existence of a homeostatic fixed point ($\nu_{FP}$). If a neuron's firing rate exceeds this target, the LTD component of the STDP rule dominates, weakening synapses and lowering the rate. If the neuron fires too slowly, LTP dominates. This creates a self-regulating loop that maintains the neuron within a functional dynamic range, effectively solving the stability problem of traditional Hebbian rules.

Implications and Future Directions

The implications of this framework extend to dendritic equalization, where "anti-STDP" rules might allow neurons to compensate for their own geometry. This can lead to a state where synapses far from the cell body (soma) become stronger to compensate for signal attenuation, ensuring all inputs have an equal "vote" regardless of their physical location.

However, the authors note several limitations. The current analysis relies on simplified linear Poisson neuron models, which may not capture the full complexity of voltage-gated dynamics in real biological neurons. Additionally, the interplay between postsynaptic expression of plasticity and presynaptic changes (such as neurotransmitter release probability) remains an area for future research. Integrating these various timescales—from millisecond spike timing to long-term homeostatic maintenance—remains a central challenge for computational neuroscience.

Summary updated

Abstract

Synaptic plasticity is thought to be the neuronal correlate of learning. Moreover, modification of synapses contributes to the activity-dependent homeostatic maintenance of neurons and neural networks. In this chapter, we review theories of synaptic plasticity and show that both homeostatic control of activity and detection of correlations in the presynaptic input can arise from spike-timing dependent plasticity (STDP). Relations to classical rate-based Hebbian learning are discussed.
BibTeX
@inproceedings{abbott2003homeostasis,
  author = {L. F. Abbott and Wulfram Gerstner},
  title = {{Homeostasis and Learning through Spike-Timing Dependent Plasticity}},
  booktitle = {Infoscience},
  year = {2003},
}
Figure