SNN Training Methods & Supporting Results

The remaining reference papers — SNN-training methods, balanced-state and efficiency results — that sit outside the two role-grouped backbones in this literature collection: ar008 (gamma & PING) and ar007 (uncertainty & Bayesian inference). A flat catch-all rather than a curated backbone. Each entry is a one-line summary; several are already load-bearing in the codebase.

The papers

Paper	Year
Sterling & Laughlin — Principles of Neural Design	2015
London, Roth, Beeren, Häusser & Latham — Sensitivity to Perturbations in vivo Implies High Noise and Suggests Rate Coding in Cortex	2010
Renart, de la Rocha, Bartho, Hollender, Parga, Reyes & Harris — The Asynchronous State in Cortical Circuits	2010
Neftci, Mostafa & Zenke — Surrogate Gradient Learning in Spiking Neural Networks	2019
Cramer, Stradmann, Schemmel & Zenke — The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Networks	2022
Eshraghian, Ward, Neftci, Wang, Lenz, Dwivedi, Bennamoun, Jeong & Lu — Training Spiking Neural Networks Using Lessons From Deep Learning	2023
Yan, Yang, Wu, Liu, Zhang, Li, Tan & Wu — Efficient and Robust Temporal Processing with Neural Oscillations Modulated Spiking Neural Networks	2025
Burghi, Pugliese Carratelli & Rule — Costate Control for Nonlinear System Identification with Applications to Excitable Systems (preprint)	2024
Parthasarathy, Burghi & O’Leary — Temporal Discretisation Shapes Optimisation Landscape in Spiking Neural Networks (preprint)	2024

Summaries

Sterling & Laughlin (2015). Book. Argues that the brain’s architecture follows a small set of efficiency principles — compute with chemistry where possible, send information at the lowest acceptable rate, minimise wire, match channel capacity to demand — that together explain why neural circuits are built the way they are. The intellectual backdrop for any “the rhythm buys per-spike economy” claim: spikes are expensive, and the cortex is under selection to use fewer of them.

London, Roth, Beeren, Häusser & Latham (2010). Shows experimentally that a single extra spike in one rat-barrel-cortex neuron produces ≈ 28 additional spikes in its postsynaptic targets — i.e. cortical networks amplify perturbations. The amplification implies high intrinsic noise and argues that cortex must use a rate code rather than precise spike timing. The in-vivo counterpart of the chaos probe in nb050: perturbations grow, so the network is chaotic and spike-timing is not reproducible.

Renart, de la Rocha, Bartho, Hollender, Parga, Reyes & Harris (2010). Shows that recurrent networks can generate an asynchronous state with arbitrarily low pairwise spiking correlations despite heavy shared input, because E and I fluctuations track each other and the resulting negative correlation in synaptic currents cancels the shared-input correlation. The conductance-based-spiking demonstration of the Vreeswijk–Sompolinsky balanced state — the paper nb050 follows to reach near-zero pairwise correlations.

Neftci, Mostafa & Zenke (2019). The canonical tutorial-review of surrogate-gradient learning: replace the non-differentiable spike threshold with a smooth pseudo-derivative on the backward pass so that BPTT can train spiking networks. The method the entire collection trains with — every COBA/PING network in ar006 uses a fast-sigmoid surrogate exactly as described here.

Cramer, Stradmann, Schemmel & Zenke (2022). Introduces the Spiking Heidelberg Digits (SHD) and Spiking Speech Commands datasets — spike-train classification benchmarks where leveraging spike timing (not just rate) is necessary for good accuracy. SHD is the collection’s canonical multi-class spiking benchmark (the shd dataset slug in the trainer).

Eshraghian, Ward, Neftci, Wang, Lenz, Dwivedi, Bennamoun, Jeong & Lu (2023). A book-length tutorial-and-perspective on training SNNs with deep-learning tools, and the paper behind the snnTorch library. The collection’s readout choices (mem-mean, spike-count) and surrogate setup follow this paper’s tutorials directly.

Yan, Yang, Wu, Liu, Zhang, Li, Tan & Wu (2025). “Rhythm-SNN”: modulating spiking neurons with heterogeneous oscillatory signals so they activate periodically at distinct frequencies, which markedly reduces firing rates while improving temporal-processing robustness and energy efficiency. Essentially the machine-learning statement of this collection’s thesis — oscillatory gating buys sparsity and efficiency — arrived at from the applied-SNN side.

Burghi, Pugliese Carratelli & Rule (preprint). Introduces Surrogate Gradients by Costate Control (SGCC): treats the exploding/vanishing costate (adjoint) dynamics of BPTT through excitable systems as a control problem, and designs gradient regularisers — a proportional high-gain controller and a state-dependent one — that provably tame the instability without changing the fixed points. The theoretical foundation for the —v-grad-dampen gradient-stabilisation flag documented in ar006.

Parthasarathy, Burghi & O’Leary (preprint). Shows that the temporal discretisation step $\Delta t$ shapes both the forward dynamics and the backward gradient flow of an SNN, so spiking patterns and surrogate-gradient learning are sensitive to $\Delta t$ — a primary contributor to the poor cross-implementation reproducibility of SNN results. Directly relevant to the $\Delta t$ -stability audits in this project (nb044): it is the methodological warning those notebooks are answering.