Analytical Solution of Spike-timing Dependent Plasticity Based on Synaptic Biophysics

Bernd Porr, Ausra Saudargiene, Florentin WörgötterNeurIPS 2003

Introduction

In neuroscience, "Spike-Timing Dependent Plasticity" (STDP) describes how the strength of a connection (synapse) between neurons changes based on the millisecond-level timing of their activity. If the presynaptic "sender" neuron fires just before the postsynaptic "receiver" neuron, the connection strengthens (Long-Term Potentiation, or LTP). If the order is reversed, the connection weakens (Long-Term Depression, or LTD).

While this phenomenon is well-documented, most existing models are either purely descriptive—fitting arbitrary curves to experimental data—or rely on computationally heavy simulations. There has long been a gap in understanding the precise biophysical "why" behind the shape of these learning curves. This paper provides a rare analytical (mathematical) derivation of STDP by modeling the underlying biophysics of a neuron, seeking a closed-form mathematical solution to explain why some parts of a neuron exhibit asymmetric STDP while others show symmetric Hebbian learning.

The Biophysical Model

The authors derive a learning rule based on two primary biophysical components interacting at the synapse:

  1. Presynaptic side: The conductance of NMDA channels, which are unique because they are sensitive to both the presence of neurotransmitters and the cell's internal voltage.
  2. Postsynaptic side: The "backpropagating (BP) spike"—an electrical signal that travels backward from the neuron's main body (soma) into its receiving branches (dendrites).
Biophysical model of a plastic synapse showing NMDA conductance and the backpropagating spike.

The "Secret Sauce": Instead of simply correlating overall activity levels, the authors hypothesize that synaptic plasticity is driven by correlating the NMDA channel conductance with the derivative (the rate of change) of the membrane potential ($dv/dt$). This change in calcium concentration, linked to $dv/dt$, proves to be a better predictor of synaptic plasticity than absolute concentration levels, naturally explaining why the learning curve has both positive and negative phases.

To solve this mathematically, the researchers simplified the membrane equations into a band-pass filter model and used a Taylor expansion to linearize the complex, non-linear NMDA conductance. By applying Plancherel’s Theorem, they solved the correlation in the Laplace (frequency) domain, resulting in a set of equations that describe weight changes as a function of the time difference ($T$) between spikes.

Key Insights: Shape Dictates Function

A major revelation of this analytical approach is that dendrites are not just passive wires; they act as active filters. As a BP-spike travels further from the soma to the distal (far) dendrites, it slows down and its shape spreads out.

Because the BP-spike changes shape as it travels, the resulting learning rule changes depending on the synapse's location:

  • Near the Soma: Where BP-spikes are sharp and fast (short rise/decay times), the math produces the classic, sharply asymmetric STDP curve required for temporal sequence learning.
  • In Distal Dendrites: Where BP-spikes are broad and slow (long decay), the equations naturally shift toward temporally symmetric Hebbian learning, which is often used for Principal Component Analysis (PCA) in neural networks.
Analytical STDP curves and corresponding BP-spike shapes, demonstrating the transition from asymmetric to symmetric learning.

Results and Implications

The derived equations successfully reproduce the bimodal STDP curves seen in biological experiments. For realistic, small currents, the first term of their derivation (representing the NMDA influence) dominates the curve, providing a highly stable fit for biological data.

Ultimately, this model presents a Unified Theory of synaptic plasticity. It suggests that asymmetric STDP and symmetric Hebbian learning are not two fundamentally different mechanisms, but rather two ends of a single biophysical continuum determined by the shape of the backpropagating spike.

Limitations and Future Work

While mathematically elegant, the model has limitations. The authors used a Taylor expansion around 0mV to simplify the highly non-linear NMDA channel, which might lose some nuances of the channel's behavior near the resting potential. Additionally, the model focuses almost exclusively on NMDA channels as the source of calcium, whereas in reality, voltage-gated calcium channels and internal stores also contribute to plasticity.

Future work will need to investigate how these varying learning rules—shifting from asymmetric to symmetric across different parts of the same dendrite—contribute to the overall computational power and information processing capabilities of a single neuron.

Summary updated

BibTeX
@inproceedings{porr2003analytical,
  author = {Bernd Porr and Ausra Saudargiene and Florentin Wörgötter},
  title = {{Analytical Solution of Spike-timing Dependent Plasticity Based on Synaptic Biophysics}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2003},
}
Figure