Poisson balanced spiking networks

Camille E. Rullán Buxó, Jonathan W. PillowbioRxiv 2019 · DOI

Abstract

An important problem in computational neuroscience is to understand how networks of spiking neurons can carry out various computations underlying behavior. Balanced spiking networks (BSNs) provide a powerful framework for implementing arbitrary linear dynamical systems in networks of integrate-and-fire neurons. However, the classic BSN model requires near-instantaneous transmission of spikes between neurons, which is biologically implausible. Introducing realistic synaptic delays leads to a pathological regime known as “ping-ponging”, in which different populations spike maximally in alternating time bins, causing network output to overshoot the target solution. Here we document this phenomenon and provide a novel solution: we show that a network can have realistic synaptic delays while maintaining accuracy and stability if neurons are endowed with conditionally Poisson firing. Formally, we propose two alternate formulations of Poisson balanced spiking networks: (1) a “local” framework, which replaces the hard integrate-and-fire spiking rule within each neuron by a “soft” threshold function, such that firing probability grows as a smooth nonlinear function of membrane potential; and (2) a “population” framework, which reformulates the BSN objective function in terms of expected spike counts over the entire population. We show that both approaches offer improved robustness, allowing for accurate implementation of network dynamics with realistic synaptic delays between neurons. Moreover, both models produce positive correlations between similarly tuned neurons, a feature of real neural populations that is not found in the original BSN. This work unifies balanced spiking networks with Poisson generalized linear models and suggests several promising avenues for future research.

Introduction: The Stability vs. Realism Trade-off

In computational neuroscience, the Balanced Spiking Network (BSN) framework is a prominent model for understanding how neurons perform efficient computation. BSNs operate on a "greedy" principle: a neuron fires only when its spike reduces the error between the network's current output and a target signal. While mathematically elegant, this framework relies on a biologically impossible assumption: instantaneous communication.

In real brains, synaptic delays—the time it takes for a signal to cross from one neuron to another—are inevitable. When these delays (even as small as 1ms) are introduced into a standard BSN, the system collapses into an unstable state called "ping-ponging."

The "ping-pong" effect: without restrictions, synaptic delays cause massive oscillations in network output as neurons over-correct for errors they cannot yet see.

As shown in Figure 4, when there is no restriction on spiking, neurons within a population fire simultaneously to correct a perceived error. Because of the delay, they don't see each other's spikes immediately, leading to a massive overshoot. An opposing population then fires to correct this new error, resulting in high-frequency, unstable oscillations that destroy the accuracy of the computation.

Methods: Introducing Poisson Stochasticity

To resolve this instability, the authors propose moving from deterministic "Integrate-and-Fire" rules to probabilistic, conditionally Poisson firing. They introduce two primary frameworks:

  1. The Local Framework: This approach replaces the "hard" spike threshold with a "soft" sigmoidal function. As a neuron's membrane potential (representing the local error) increases, its probability of firing increases smoothly.
  2. The Population Framework: This top-down approach calculates a vector of firing rates ($\lambda$) that minimizes the expected error across the entire population, distributing the computational load more gracefully.
The Local Poisson BSN mechanism: replacing hard thresholds with a sigmoidal spike rate function (A-C) introduces jitter that prevents synchronized "ping-pong" spiking (D-E).

As illustrated in Figure 6, the sigmoidal threshold (A) allows for a range of firing rates. By adjusting the steepness ($\alpha$), the network can transition from highly stochastic firing to a regime that approximates the original BSN but remains stable under delay. This stochasticity breaks the symmetry that causes neurons to fire in lockstep, effectively "smoothing" the network's collective response.

Key Insights and Results

The transition to Poisson spiking provides several critical advantages:

  • Robustness to Delay: The Poisson models maintain high accuracy ($R^2 \approx 0.9$) even with synaptic delays of 2–4ms, whereas original BSNs fail almost immediately.
  • Biological Plausibility: A common critique of BSNs is that they produce negative correlations between similarly tuned neurons (due to competition). Real neural data shows positive correlations. The Poisson framework naturally replicates these positive correlations, aligning the model with experimental observations.
  • Theoretical Unification: This work bridges the gap between BSNs and Generalized Linear Models (GLMs), which are standard tools for analyzing experimental spike trains. It suggests that the brain's "noise" might actually be a feature that provides stability to its underlying dynamical computations.

Conclusion

By incorporating uncertainty into the spiking process, the authors have created a version of Balanced Spiking Networks that is both computationally powerful and biologically realistic. This framework eliminates the pathological oscillations caused by synaptic delays and provides a more accurate statistical description of real neural populations. Future work will look to extend these principles to nonlinear dynamical systems and explore how these networks might learn their connectivity through plasticity rules.

Summary updated

Abstract

Abstract An important problem in computational neuroscience is to understand how networks of spiking neurons can carry out various computations underlying behavior. Balanced spiking networks (BSNs) provide a powerful framework for implementing arbitrary linear dynamical systems in networks of integrate-and-fire neurons (Boerlin et al. [1]). However, the classic BSN model requires near-instantaneous transmission of spikes between neurons, which is biologically implausible. Introducing realistic synaptic delays leads to an pathological regime known as “ping-ponging”, in which different populations spike maximally in alternating time bins, causing network output to overshoot the target solution. Here we document this phenomenon and provide a novel solution: we show that a network can have realistic synaptic delays while maintaining accuracy and stability if neurons are endowed with conditionally Poisson firing. Formally, we propose two alternate formulations of Poisson balanced spiking networks: (1) a “local” framework, which replaces the hard integrate-and-fire spiking rule within each neuron by a “soft” threshold function, such that firing probability grows as a smooth nonlinear function of membrane potential; and (2) a “population” framework, which reformulates the BSN objective function in terms of expected spike counts over the entire population. We show that both approaches offer improved robustness, allowing for accurate implementation of network dynamics with realistic synaptic delays between neurons. Moreover, both models produce positive correlations between similarly tuned neurons, a feature of real neural populations that is not found in the original BSN. This work unifies balanced spiking networks with Poisson generalized linear models and suggests several promising avenues for future research.
BibTeX
@misc{buxo2019poisson,
  author = {Camille E. Rullán Buxó and Jonathan W. Pillow},
  title = {{Poisson balanced spiking networks}},
  year = {2019},
  doi = {10.1101/836601},
  howpublished = {bioRxiv},
}
Figure