A computational theory of spike-timing dependent plasticity: achieving robust neural responses via conditional entropy minimization.

Sander M. Bohté, Michael C. MozerCentrum 2005

Abstract: Experimental studies have observed synaptic potentiation when a presynaptic neuron fires shortly before a postsynaptic neuron, and synaptic depression when the presynaptic neuron fires shortly after. The dependence of synaptic modulation on the precise timing of the two action potentials is known as spike-timing dependent plasticity or STDP. We derive STDP from a simple computational principle: synapses adapt so as to minimize the postsynaptic neuron's variability to a given presynaptic input, causing the neuron's output to become more reliable in the face of noise. Using an entropy-minimization objective function and the biophysically realistic spike-response model of Gerstner (2001), we simulate neurophysiological experiments and obtain the characteristic STDP curve along with other phenomena including the reduction in synaptic plasticity as synaptic efficacy increases. We compare our account to other efforts to derive STDP from computational principles, and argue that our account provides the most comprehensive coverage of the phenomena. Thus, reliability of neural response in the face of noise may be a key goal of unsupervised cortical adaptation.

Introduction

In the field of neuroscience, Spike-Timing Dependent Plasticity (STDP) represents a cornerstone of our understanding of synaptic learning. Discovered in the late 1990s, STDP describes how the strength of a synapse changes based on the millisecond-level timing between presynaptic and postsynaptic spikes. Specifically, Long-Term Potentiation (LTP) occurs when a presynaptic neuron fires just before the postsynaptic target, while Long-Term Depression (LTD) occurs when it fires just after.

While STDP is widely accepted as a biological fact, the functional "why" behind this specific timing dependence has remained a subject of intense debate. This report from CWI (2005) by Sander Bohte and Michael Mozer proposes that STDP is not merely a variant of Hebbian learning, but rather an emergent property of a more fundamental computational goal: conditional entropy minimization.

The Principle of Reliability

The authors hypothesize that the primary objective of synaptic adaptation is to ensure that a neuron produces a reliable, consistent output in response to a specific input pattern, despite the inherent stochasticity (noise) of biological systems. By minimizing the variability—or entropy—of the postsynaptic response, a neuron becomes a more effective and predictable information processor.

In this framework, STDP is viewed as a mechanism for noise suppression. When a presynaptic spike arrives just before a postsynaptic spike, it is a "useful" signal that likely contributed to the output; strengthening this connection (LTP) makes that output more certain in the future. Conversely, a spike arriving just after the postsynaptic response is redundant or "noisy." Weakening this connection (LTD) prevents that late input from potentially triggering an unwanted, secondary spike, thereby reducing output variability.

Modeling and Results

To test this theory, the authors employed a Stochastic Spike Response Model (sSRM). This model is more biophysically grounded than standard integrate-and-fire models, as it accounts for the refractory period—the time during which a neuron is less likely to fire again after an initial spike.

Figure 1: (a) Schematic of the pre- and postsynaptic interaction. (b) The characteristic STDP curve showing LTP and LTD as a function of spike timing.

By applying gradient descent to an objective function representing the conditional entropy of the postsynaptic response, the authors derived a learning rule. As shown in Figure 1, this derived rule produces a weight-change curve that closely matches the experimental data observed in landmark studies such as Zhang et al. (1998) and Bi & Poo (1998). The model successfully captures both the LTP and LTD phases of the STDP window.

Key Insights

Beyond fitting the basic STDP curve, the entropy-minimization approach explains several other complex biological phenomena:

  • Natural Saturation: In biological systems, as a synapse grows stronger, it becomes harder to potentiate further. This model predicts this "saturation" naturally: as a synapse becomes highly effective, the neuron's response becomes nearly certain, the entropy gradient approaches zero, and learning slows down.
  • LTD as Error Correction: The model provides a functional reason for the weakening of synapses. If a late pulse is perceived as noise that might have triggered an unwanted second spike, the neuron weakens that connection to ensure that "extra" spike doesn't happen next time.
  • The "LTD Dip": The model predicts a subtle weakening of the synapse even at very small negative time intervals (where the pre-spike is very close to the post-spike), a phenomenon that has been observed in specific biological experiments (e.g., Nishiyama et al., 2000).
  • Parameter Sensitivity: The width of the STDP window (typically 20-30ms) is shown to be directly linked to the physical time constants of the neuron's membrane, bridging the gap between low-level physics and high-level learning rules.

Conclusion

Bohte and Mozer demonstrate that the complex behavior of STDP can be derived from the simple, elegant principle of making neural responses more reliable. While the mathematical integration required for entropy minimization is complex, the authors suggest that biological STDP may be an "evolved shortcut"—a simple rule that neurons follow to achieve the sophisticated goal of noise-robust information processing. This perspective shifts the view of unsupervised cortical adaptation from simple association to a rigorous process of reliability optimization.

Summary updated

Abstract

Experimental studies have observed synaptic potentiation when a presynaptic neuron fires shortly before a postsynaptic neuron, and synaptic depression when the presynaptic neuron fires shortly after. The dependence of synaptic modulation on the precise timing of the two action potentials is known as spike-timing dependent plasticity or STDP. We derive STDP from a simple computational principle: synapses adapt so as to minimize the postsynaptic neuron's variability to a given presynaptic input, causing the neuron's output to become more reliable in the face of noise. Using an entropy-minimization objective function and the biophysically realistic spike-response model of Gerstner (2001), we simulate neurophysiological experiments and obtain the characteristic STDP curve along with other phenomena including the reduction in synaptic plasticity as synaptic efficacy increases. We compare our account to other efforts to derive STDP from computational principles, and argue that our account provides the most comprehensive coverage of the phenomena. Thus, reliability of neural response in the face of noise may be a key goal of unsupervised cortical adaptation
BibTeX
@inproceedings{bohte2005computational,
  author = {Sander M. Bohté and Michael C. Mozer},
  title = {{A computational theory of spike-timing dependent plasticity: achieving robust neural responses via conditional entropy minimization.}},
  booktitle = {Centrum},
  year = {2005},
}
Figure