arxiv: 2605.11835 · v1 · submitted 2026-05-12 · 💻 cs.NE · cs.AI· cs.LG

Recognition: no theorem link

Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing

Alex Fulleda-Garcia , Saray Soldado-Magraner , Josep Maria Margarit-Taul\'e

Authors on Pith no claims yet

Pith reviewed 2026-05-13 05:07 UTC · model grok-4.3

classification 💻 cs.NE cs.AIcs.LG

keywords spiking neural networksconductance-based neuronsmulti-timescale dynamicsdirect backpropagationtemporal regressionMackey-Glassactivity sparsityneuromorphic computing

0 comments

The pith

Multi-timescale conductance spiking networks support direct backpropagation and deliver higher accuracy with far fewer spikes than standard models on time-series tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces spiking neurons whose dynamics arise from adjusting conductances at three distinct timescales to shape the current-voltage curve. This single parametrization produces multiple firing patterns while remaining fully differentiable, allowing a discrete-time implementation that supports exact backpropagation through time. When tested on Mackey-Glass time-series regression, the resulting networks exceed the accuracy of both LIF and AdLIF baselines yet generate substantially lower spiking activity. The approach therefore addresses the usual trade-off between trainability, dynamical richness, and sparsity in spiking models for temporal processing.

Core claim

Parametrizing the I-V curve with fast, slow, and ultra-slow conductances produces rich firing regimes including tonic, phasic, and bursting responses within one neuron model. The resulting dynamics admit an exact discrete-time formulation that permits direct backpropagation through time without surrogate gradients. On Mackey-Glass regression, feedforward networks built from these neurons outperform LIF and AdLIF models while exhibiting markedly sparser activity from both communication and computation standpoints.

What carries the argument

Multi-timescale conductance parametrization of the current-voltage curve, which sets excitability and yields diverse firing regimes while remaining efficiently differentiable.

If this is right

Networks of these neurons can be trained end-to-end with standard backpropagation through time.
A single neuron model can generate tonic, phasic, and bursting responses without changing architecture.
Lower overall spiking rates reduce both communication bandwidth and computational cost in temporal tasks.
The continuous dynamics map directly to efficient analog circuit implementations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same conductance shaping could be applied to recurrent layers to improve long-horizon sequence modeling without added surrogate tricks.
Matching the three timescales to the dominant frequencies of a given dataset might further reduce required network size on real-world signals.
Because the model already supports analog-circuit mapping, it could serve as a drop-in block for mixed-signal neuromorphic chips targeting always-on temporal sensing.

Load-bearing premise

That adjusting the three conductance timescales simultaneously supplies systematic control over firing regimes, preserves differentiability, and stays computationally cheap enough to outperform simpler models on regression tasks.

What would settle it

Running the identical Mackey-Glass regression experiments and finding that the multi-timescale networks fail to exceed LIF or AdLIF accuracy or show no reduction in spike count.

Figures

Figures reproduced from arXiv: 2605.11835 by Alex Fulleda-Garcia, Josep Maria Margarit-Taul\'e, Saray Soldado-Magraner.

**Figure 2.** Figure 2: Diversity of neuronal firing patterns. The figure illustrates four distinct firing regimes generated by the neuron model in response to external stimulation. In each panel, the upper trace represents the membrane potential, Um, and the lower trace indicates the input current, Iin(t). The patterns include Tonic Spiking and Tonic Bursting (sustained responses to constant input), Phasic Spiking and Phasic Bur… view at source ↗

read the original abstract

Spiking neural networks (SNNs) promise low-power event-driven computation for temporally rich tasks, but commonly used neuron models often trade off gradient-based trainability, dynamical richness, and high activity sparsity. These limitations are acute in regression, where approximation error, noise and spike discretization can severely degrade continuous-valued outputs. Indeed, many state-of-the-art (SOTA) SNNs rely on simple phenomenological dynamics trained with surrogate gradients and offer limited control over spiking diversity and sparsity. To overcome such limitations, we introduce multi-timescale conductance spiking networks, a gradient-trainable framework in which neural dynamics emerge from shaping the current-voltage (I-V) curve by tuning fast, slow and ultra-slow conductances. This parametrization allows systematic control over excitability, can be implemented efficiently in analog circuits, and yields rich firing regimes including tonic, phasic and bursting responses within a single model. We derive a discrete-time formulation of these differentiable dynamics, enabling direct backpropagation through time without surrogate-gradient approximations. To probe both trainability and accuracy, we evaluate feedforward networks of these neurons at the predictability limit of Mackey-Glass time-series regression and compare them to baseline LIF and SOTA AdLIF networks. Our model outperforms LIF and AdLIF networks, while exhibiting substantially sparser activity from both communication and computational perspectives. These results highlight multi-timescale conductance spiking neurons as a promising building block for energy-aware temporal processing and neuromorphic implementation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a multi-timescale conductance neuron model that supports direct BPTT without surrogates and shows lower error plus sparser activity than LIF or AdLIF on Mackey-Glass regression.

read the letter

The core contribution is a spiking neuron whose dynamics come from tuning fast, slow, and ultra-slow conductances to shape the I-V curve. This single parametrization produces multiple firing regimes and stays fully differentiable after Euler discretization, so training uses plain backpropagation through time. They report better regression accuracy on Mackey-Glass together with lower spike rates and fewer arithmetic operations per step than the LIF and AdLIF baselines. The update rules are written out explicitly and the approach avoids surrogate gradients by construction. That combination of biophysical motivation, direct trainability, and measured sparsity is the part worth paying attention to. The results line up with the claim that the model gives richer dynamics while remaining efficient for temporal tasks. Evaluation stays on one benchmark, which limits how far the superiority claim can be pushed without more datasets or controls for network size and hyperparameter search. No error bars or statistical tests are mentioned in the summary, though the stress-test note indicates the full text supplies the training protocol and raw numbers. The work is aimed at people building SNNs for neuromorphic or low-power temporal processing. It is worth a serious referee because the equations are concrete, the differentiability argument holds, and the sparsity numbers are a practical plus even if further validation is needed.

Referee Report

1 major / 1 minor

Summary. The paper introduces multi-timescale conductance spiking networks (MCSNs) in which neuron dynamics emerge from parametrizing the I-V curve with fast, slow, and ultra-slow conductances. It derives an explicit Euler-discretized discrete-time formulation that remains fully differentiable, enabling direct backpropagation through time without surrogate gradients. Feedforward networks built from these neurons are evaluated on Mackey-Glass time-series regression and claimed to outperform both LIF and AdLIF baselines while producing substantially sparser spiking activity from communication and computational standpoints.

Significance. If the reported gains hold under rigorous statistical scrutiny, the work would provide a useful addition to the SNN literature by supplying a biophysically motivated neuron model that combines direct gradient trainability, controllable dynamical richness (tonic, phasic, bursting), and inherent sparsity. The explicit derivation of the update rules and the avoidance of surrogate-gradient approximations constitute clear methodological strengths that improve reproducibility of the core contribution.

major comments (1)

[Experimental evaluation section] In the experimental evaluation of Mackey-Glass regression, the manuscript reports lower MSE together with reduced spike counts and arithmetic operations relative to LIF and AdLIF, yet supplies neither error bars, exact network sizes, training hyperparameters, nor statistical significance tests. These omissions prevent a reader from determining whether the claimed superiority is robust or sensitive to implementation details.

minor comments (1)

[Abstract] The abstract states that the model 'outperforms LIF and AdLIF networks' and exhibits 'substantially sparser activity' but does not include the numerical deltas in MSE or spike-rate reduction; adding these quantities would make the summary more informative.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed and constructive review. We agree that the experimental section requires additional rigor to substantiate the reported performance gains and have revised the manuscript to address this concern.

read point-by-point responses

Referee: In the experimental evaluation of Mackey-Glass regression, the manuscript reports lower MSE together with reduced spike counts and arithmetic operations relative to LIF and AdLIF, yet supplies neither error bars, exact network sizes, training hyperparameters, nor statistical significance tests. These omissions prevent a reader from determining whether the claimed superiority is robust or sensitive to implementation details.

Authors: We thank the referee for this observation. The original manuscript indeed omitted these elements, which limits interpretability. In the revised version, we now report mean MSE and spike counts with error bars (standard deviation over 10 independent runs using different random seeds), explicitly list the network architectures (e.g., identical hidden-layer sizes of 128 neurons for all models), provide a complete hyperparameter table (learning rate, optimizer, batch size, number of epochs, and discretization timestep), and include statistical significance testing via paired t-tests showing p < 0.01 for the MSE improvements over both baselines. These additions confirm the robustness of the results under the reported conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper defines multi-timescale conductance spiking neurons from biophysical shaping of the I-V curve via fast/slow/ultra-slow conductances, then applies standard Euler discretization to obtain explicit, everywhere-differentiable update rules that enable direct BPTT. These equations are presented as derived constructs rather than fitted to the target task; performance (lower MSE, lower spike counts, lower arithmetic operations) is measured on external Mackey-Glass regression against independent LIF/AdLIF baselines. No load-bearing step equates a claimed prediction to its own inputs by construction, no self-citation chain justifies uniqueness, and no ansatz is smuggled in; the central claims remain falsifiable against external data and standard neuron models.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard conductance-based neuron assumptions and the existence of tunable fast/slow/ultra-slow conductances as free parameters; no new physical entities are postulated.

free parameters (1)

fast, slow, and ultra-slow conductance parameters
Tuned to shape the I-V curve and produce desired excitability and firing regimes; values are not specified in abstract.

axioms (2)

domain assumption Conductance-based neuron dynamics can be discretized while preserving differentiability for direct BPTT
Invoked to enable gradient training without surrogates.
domain assumption Standard biophysical I-V curve shaping principles apply to the multi-timescale model
Underlies the claim of rich firing regimes within a single model.

pith-pipeline@v0.9.0 · 5591 in / 1401 out tokens · 66191 ms · 2026-05-13T05:07:49.637340+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

[1]

semi-digital

approximating the backward pass derivative with a smoothed function (e.g., a sigmoid or arctan). Although this combination has enabled deep SNN train- ing and competitive performance on standard benchmarks, it comes with several limitations that become especially pronounced for continuous-valued temporal regression. First, LIF-based neurons strip away the...

work page
[2]

The series is generated using the standard parameters γ = 0 .1, β = 0 .2, n = 10 and a delay τ = 17 , which places the system in a chaotic regime

Dataset: We employ the Mackey–Glass (MG) chaotic time series [27], a standard benchmark for reservoir computing and neuromorphic forecasting tasks [30], [31]. The series is generated using the standard parameters γ = 0 .1, β = 0 .2, n = 10 and a delay τ = 17 , which places the system in a chaotic regime. dx(t) dt = β x(t − τ ) 1 + x(t − τ )n − γx(t) (10) ...

work page
[3]

The Leaky Integrate- and-Fire serves as the standard non-adaptive baseline

Comparative Baselines: To isolate the performance ben- eﬁts of the proposed conductance-based dynamics, we bench- mark our model against two widely used spiking neuron formulations: the LIF and AdLIF models. The Leaky Integrate- and-Fire serves as the standard non-adaptive baseline. To ensure a robust implementation, we applied the optimized modules provi...

work page
[4]

Network architecture: All models utilize a feed-forward spiking architecture with no lateral or recurrent synaptic connections. The topology consists of four stages: a linear input projection ( 1×N ) that maps the scalar time series into a high-dimensional space; a hidden processing layer containing N independent spiking neurons; a readout layer ( N × 1) ...

work page
[5]

fair comparison

Training details and Hyperparameters: To ensure a rigorous and reproducible evaluation, we adopted a structured multi-stage optimization protocol. The training con ﬁguration encompasses the optimization framework, gradient estimation strategies, weight initialization, and a hierarchical hyperparam- eter search. a) Optimization Framework: All models were t...

work page
[6]

Neuron Dynamics Tuning: For all models, we per- formed a grid search over key intrinsic parameters (decay rates (LIF); adaptation time constants and feed- forward gain (AdLIF); and conductance time constants (MTC)) to minimize validation error. For our MTC model, other parameters were manually tuned by an- alyzing the phase-space dynamics, I–V curves and ...

work page
[7]

Architecture and Global Parameters: A random search was conducted over the macro-parameters: input window size ( Tx), dataset size ( Nsamples), hidden layer dimension ( N ), and base learning rate

work page
[8]

Cross-compensation of FET sensor drift and matrix effects in the industrial continuous monitoring of ion concentra- tions,

J. M. Margarit-Taulé, M. Martín-Ezquerra, R. Escudé-Pujol, C. Jiménez- Jorquera, and S.-C. Liu, “Cross-compensation of FET sensor drift and matrix effects in the industrial continuous monitoring of ion concentra- tions,” Sensors and Actuators B: Chemical , vol. 353, p. 131123, 2022

work page 2022
[9]

Remaining useful life estimation in prognostics using deep convolution neural networks,

X. Li, Q. Ding, and J.-Q. Sun, “Remaining useful life estimation in prognostics using deep convolution neural networks,” Reliability Engineering & System Safety , vol. 172, pp. 1–11, 2018

work page 2018
[10]

Multisensing wearables for real-time monitoring of sweat electrolyte biomarkers during exercise and analysis on their correlation with core body temperature,

S. Wang, M. Rovira, S. Demuru, C. Lafaye, J. Kim, B. P . Kunnel, C. Besson, C. Fernandez-Sanchez, F. Serra-Graells, J. M. Margarit- Taulé, J. Aymerich, J. Cuenca, I. Kiselev, V . Gremeaux, M. Saubade, C. Jimenez-Jorquera, D. Briand, and S.-C. Liu, “Multisensing wearables for real-time monitoring of sweat electrolyte biomarkers during exercise and analysis...

work page 2023
[11]

Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

T. Haarnoja, A. Zhou, P . Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” in International Conference on Machine Learning . PMLR, 2018, pp. 1861–1870

work page 2018
[12]

Training spiking neural networks using lessons from deep learning,

J. K. Eshraghian, M. Ward, E. Neftci, X. Wang, G. Lenz, G. Dwivedi, D. S. Bennamoun, Mohammed andhee, and W. D. Lu, “Training spiking neural networks using lessons from deep learning,” Proceedings of the IEEE, vol. 111, no. 9, pp. 1016–1054, 2023

work page 2023
[13]

Spatio-temporal backpropa- gation for training high-performance spiking neural networks,

Y . Wu, L. Deng, G. Li, J. Zhu, and L. Shi, “Spatio-temporal backpropa- gation for training high-performance spiking neural networks,” Frontiers in Neuroscience , vol. 12, p. 331, 2018

work page 2018
[14]

Superspike: Surrogate gradient learning in spiking neural networks,

F. Zenke and S. Ganguli, “Superspike: Surrogate gradient learning in spiking neural networks,” Neural Computation, vol. 30, no. 6, pp. 1514– 1541, 2018

work page 2018
[15]

Which model to use for cortical spiking neurons?

E. M. Izhikevich, “Which model to use for cortical spiking neurons?” IEEE Transactions on Neural Networks , vol. 15, no. 5, pp. 1063–1070, 2004

work page 2004
[16]

Gerstner, W

W. Gerstner, W. M. Kistler, R. Naud, and L. Paninski, Neuronal dynamics: From single neurons to networks and models of cognition . Cambridge University Press, 2014

work page 2014
[17]

Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,

E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,” IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

work page 2019
[18]

Slayer: Spike layer error reassignment in time,

S. B. Shrestha and G. Orchard, “Slayer: Spike layer error reassignment in time,” in Advances in Neural Information Processing Systems , vol. 31, 2018

work page 2018
[19]

Long short-term memory and learning-to-learn in networks of spiking neurons,

G. Bellec, D. Salaj, A. Subramoney, R. Legenstein, and W. Maass, “Long short-term memory and learning-to-learn in networks of spiking neurons,” Advances in neural information processing systems , vol. 31, 2018

work page 2018
[20]

Spike-frequency adaptation supports network computations on temporally dispersed information,

D. Salaj, A. Subramoney, C. Kraisnikovic, G. Bellec, R. Legenstein, and W. Maass, “Spike-frequency adaptation supports network computations on temporally dispersed information,” eLife, vol. 10, p. e65459, 2021

work page 2021
[21]

Accurate and ef ﬁcient time-domain classiﬁcation with adaptive spiking recurrent neural networks,

B. Yin, F. Corradi, and S. M. Bohté, “Accurate and ef ﬁcient time-domain classiﬁcation with adaptive spiking recurrent neural networks,” Nature Machine Intelligence , vol. 3, no. 10, pp. 905–913, 2021

work page 2021
[22]

Spike frequency adaptation: bridging neural models and neuromorphic applications,

C. Ganguly, S. S. Bezugam, E. Abs, M. Payvand, S. Dey, and M. Suri, “Spike frequency adaptation: bridging neural models and neuromorphic applications,” Communications Engineering , vol. 3, p. 22, 2024

work page 2024
[23]

Advancing spatio-temporal processing through adaptation in spiking neural net- works,

M. Baronig, R. Ferrand, S. Sabathiel, and R. Legenstein, “Advancing spatio-temporal processing through adaptation in spiking neural net- works,” Nature Communications , vol. 16, no. 1, p. 5776, 2025

work page 2025
[24]

Intrinsic ﬁring patterns of diverse neocortical neurons,

B. W. Connors and M. J. Gutnick, “Intrinsic ﬁring patterns of diverse neocortical neurons,” Trends in neurosciences , vol. 13, no. 3, pp. 99– 104, 1990

work page 1990
[25]

Petilla terminology: nomenclature of features of gabaergic interneurons of the cerebral cortex,

“Petilla terminology: nomenclature of features of gabaergic interneurons of the cerebral cortex,” Nature Reviews Neuroscience , vol. 9, no. 7, pp. 557–568, 2008

work page 2008
[26]

Conditioning by subthreshold synaptic input changes the intrinsic ﬁring pattern of ca3 hippocampal neurons,

S. Soldado-Magraner, F. Brandalise, S. Honnuraiah, M. Pfeiffer, M. Moulinier, U. Gerber, and R. Douglas, “Conditioning by subthreshold synaptic input changes the intrinsic ﬁring pattern of ca3 hippocampal neurons,” Journal of neurophysiology, vol. 123, no. 1, pp. 90–106, 2020

work page 2020
[27]

Plasticity of intrinsic neuronal excitability,

D. Debanne, Y . Inglebert, and M. Russier, “Plasticity of intrinsic neuronal excitability,” Current opinion in neurobiology , vol. 54, pp. 73– 82, 2019

work page 2019
[28]

Neuromodulation of neuromorphic circuits,

L. Ribar and R. Sepulchre, “Neuromodulation of neuromorphic circuits,” IEEE Transactions on Circuits and Systems I: Regular Papers , vol. 66, no. 8, pp. 3028–3040, 2019

work page 2019
[29]

A Neuromodulable Current-Mode Silicon Neuron for Robust and Adaptive Neuromorphic Systems

L. Mendolia, C. Wen, E. Chicca, G. Indiveri, R. Sepulchre, J.-M. Redouté, and A. Franci, “A neuromodulable current-mode silicon neu- ron for robust and adaptive neuromorphic systems,” arXiv preprint arXiv:2512.01133, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[30]

A silicon neuron,

M. Mahowald and R. Douglas, “A silicon neuron,” Nature, vol. 354, no. 6354, pp. 515–518, 1991

work page 1991
[31]

Neuromorphic silicon neuron circuits,

G. Indiveri, B. Linares-Barranco, T. J. Hamilton, A. van Schaik, R. Etienne-Cummings, T. Delbruck, S.-C. Liu, P . Dudek, P . Hä ﬂiger, S. Renaud, J. Schemmel, G. Cauwenberghs, J. Arthur, K. Hynna, F. Folowosele, S. Saïghi, T. Serrano-Gotarredona, J. Wijekoon, Y . Wang, and K. Boahen, “Neuromorphic silicon neuron circuits,” Frontiers in Neuroscience, vol. ...

work page 2011
[32]

Oscillation and chaos in physiological con- trol systems,

M. C. Mackey and L. Glass, “Oscillation and chaos in physiological con- trol systems,” Science, vol. 197, no. 4300, pp. 287–289, 1977. [Online]. Available: https://www.science.org/doi/abs/10.1126/science.267326

work page doi:10.1126/science.267326 1977
[33]

Exploiting neuro-inspired dynamic sparsity for energy-ef ﬁcient intelligent perception,

S. Zhou, C. Gao, T. Delbruck, M. V erhelst, and S.-C. Liu, “Exploiting neuro-inspired dynamic sparsity for energy-ef ﬁcient intelligent perception,” Nature Communications, vol. 16, no. 1, p. 9928,

work page
[34]

Available: https://www.nature.com/articles/s41467-025- 65387-7

[Online]. Available: https://www.nature.com/articles/s41467-025- 65387-7

work page
[35]

snntorch: Deep and online learning with spiking neural networks in python,

J. K. Eshraghian, “snntorch: Deep and online learning with spiking neural networks in python,” 2023, gitHub repository. [Online]. Available: https://github.com/jeshraghian/snnTorch

work page 2023
[36]

Neu- romorphic on-chip reservoir computing with spiking neural network architectures,

S. Karki, D. Chavez Arana, A. Sornborger, and F. Caravelli, “Neu- romorphic on-chip reservoir computing with spiking neural network architectures,” arXiv preprint arXiv:2407.20547, 2024

work page arXiv 2024
[37]

Methodology based on spiking neural net- works for univariate time-series forecasting,

S. Lucas and E. Portillo, “Methodology based on spiking neural net- works for univariate time-series forecasting,” Neural Networks, vol. 173, p. 106171, 2024

work page 2024
[38]

Predicting chaotic time series,

J. D. Farmer and J. J. Sidorowich, “Predicting chaotic time series,” Phys. Rev. Lett. , vol. 59, pp. 845–848, Aug 1987. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevLett.59.845

work page doi:10.1103/physrevlett.59.845 1987
[39]

SE-adLIF: Code for

S. Baronig, M. Schöne, A. Anand, D. Kappel, G. Parton, S. Bilgic, and R. Legenstein, “SE-adLIF: Code for "advancing spatio-temporal processing in spiking neural networks through adaptation",” 2025, gitHub repository. [Online]. Available: https://github.com/IGITUGraz/SE-adlif

work page 2025