Supervised learning with decision margins in pools of spiking neurons

Charlotte Le Mouel, Kenneth D. Harris, Pierre YgerJ. Comput. Neurosci 2014 · DOI

Introduction

Categorization is a fundamental challenge for both biological and artificial neural systems. While classical models like the Perceptron can learn linear boundaries, they often lack robustness when faced with noisy or non-linearly separable data. In machine learning, Support Vector Machines (SVMs) addressed this by introducing a "decision margin"—a buffer zone that ensures classes are separated by a specific distance, rather than just a bare minimum boundary. This study explores how spiking neurons can implement similar margin-based learning principles through biologically plausible mechanisms.

Methods

The researchers proposed a learning framework for conductance-based integrate-and-fire neurons using an approximation of the SVM hinge-loss function.

Architecture and Cost Function

The model utilizes "pools" of neurons rather than isolated units. A global error signal is calculated based on the collective spike count of the pool. The learning rule is governed by two thresholds: $\theta_+$, the minimum spikes required for a target category, and $\theta_-$, the maximum spikes allowed for a null category. The region between these thresholds defines the decision margin $M = \theta_+ - \theta_-$.

Architecture and Cost Function: (A) Input generation from rate templates. (B) Pool architecture with global error signal. (C) Hinge-loss cost function defining the decision margin between thresholds.

3-Factor Learning Rule

Synaptic plasticity is driven by a three-factor rule that combines local and global information:

  1. Pre-synaptic activity: The timing of input spikes.
  2. Post-synaptic state: The neuron's membrane potential.
  3. Global error signal: The discrepancy between the pool's actual spike count and the desired margin-constrained count.

An eligibility trace at each synapse maintains a "memory" of recent activity, allowing weight updates to be applied only when the global error signal is received at the end of a trial.

Results

The framework was tested on synthetic datasets where input patterns were converted into 100ms Poisson spike trains. As the number of patterns increases relative to input dimensions, the task becomes highly linearly non-separable.

Learning Dynamics

During training, synaptic weights evolve until the pool's output satisfies the margin requirements. As shown in Figure 3, the average number of spikes emitted per pattern converges toward the defined thresholds, leading to a steady increase in classification performance. This rule is intrinsically stable because plasticity stops once the margin requirements are met.

Learning Dynamics: (A) Evolution of synaptic weights for neurons A and B. (B) Convergence of spike counts toward the thresholds. (C) Classification performance reaching near-perfect levels over training presentations.

Performance and Capacity

The study found that single neurons trained with this rule perform similarly to standard linear SVMs. However, pools of neurons exhibit superior capabilities. On difficult, non-separable tasks, pools of 10 neurons increased the network's capacity from 14 patterns (single neuron) to 34 patterns. Furthermore, the global error signal encouraged "spontaneous specialization," where individual neurons in the pool learned to respond to specific subsets of patterns, resulting in a sparse coding scheme.

Discussion

This margin-based approach provides a robust mechanism for supervised learning in spiking networks. By requesting a margin ($M > 0$), the system avoids "lazy" learning and improves generalization on non-separable problems.

While the model is computationally effective, several biological questions remain. The identity of the global error signal could correspond to neuromodulators like dopamine or specific cerebellar pathways. Future work may address the need for feedforward inhibition to handle varying input intensities and explore whether this approach can be adapted to classify complex spatiotemporal spike sequences beyond rate-based patterns.

Summary updated

BibTeX
@article{mouel2014supervised,
  author = {Charlotte Le Mouel and Kenneth D. Harris and Pierre Yger},
  title = {{Supervised learning with decision margins in pools of spiking neurons}},
  journal = {J. Comput. Neurosci.},
  year = {2014},
  doi = {10.1007/s10827-014-0505-9},
}
Figure