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

Published 12 May 2026 in cs.NE, cs.AI, and cs.LG | (2605.11835v1)

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.

Abstract PDF Upgrade to Chat

Authors (3)

Summary

The paper proposes a multi-timescale conductance spiking neuron model that enhances temporal processing with inherent gradient-trainability.
It demonstrates superior forecasting accuracy and efficiency on chaotic time series, outperforming conventional LIF and AdLIF models.
The framework integrates biophysical principles with sparse, low-power operation, paving the way for improved neuromorphic and temporal machine learning applications.

Multi-Timescale Conductance Spiking Networks: Framework and Contributions

The paper "Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing" (2605.11835) proposes a novel, conductance-based neuronal architecture for spiking neural networks (SNNs) that realizes richer intrinsic neural dynamics and enhanced temporal processing compared to conventional models. The central development is a multi-timescale conductance (MTC) spiking neuron model, directly motivated by biophysical principles, which systematically shapes neural excitability by tuning fast, slow, and ultra-slow conductances. This paradigm stands in contrast to widely used phenomenological SNN models (e.g., LIF, AdLIF), offering differentiable, gradient-trainable dynamics without resorting to surrogate gradient approximations for backpropagation through time (BPTT).

Motivation: Limitations of Conventional SNN Models

Restrictive dynamics typical of LIF and related adaptive-spiking models impose critical limitations on regression and temporally rich tasks. Phenomenological SNNs tend to optimize for computational tractability, often sacrificing internal richness and biological fidelity. The use of surrogate gradients introduces discrepancies between forward and backward computational passes, further limiting the precise learning of temporal dependencies. In these models, sparsity often emerges as a byproduct of threshold manipulation or additional regularization rather than as a direct result of interpretable, configurable neuronal dynamics.

By contrast, biological neurons exhibit dynamical diversity attributable to ionic conductances operating over disparate timescales, leading to a continuum of firing behaviors essential for temporally extended computations. This diversity, and its implication for energy efficiency and sparse computation, is largely absent from state-of-the-art SNN models.

MTC Model: Conductance-Driven, Differentiable Spiking

The MTC framework extends the compact conductance-based model of Ribar and Sepulchre into a discrete-time, machine-learning-suitable formalism. Each neuron integrates a parallel set of voltage-dependent conductance elements, each with individually tunable timescales and gains, shaping the aggregate current-voltage (I-V) curve of the neuron. The model admits arbitrary combinations of fast-negative, slow-positive, ultra-slow conductance elements, producing a controllable repertoire of firing regimes: tonic spiking, phasic spiking, and various bursting modes, among others.

Unlike LIF/AdLIF neurons, MTC neurons are inherently differentiable—no explicit non-differentiable hard thresholds are present in the core dynamics—allowing direct use of BPTT for learning. Information transfer utilizes a semi-digital, normalized saturated ReLU mapping of the membrane potential, providing a sparse, stable signaling channel that supports hardware compatibility.

Experimental Evaluation: Mackey-Glass Time Series Forecasting

The efficacy of MTC SNNs is demonstrated on chaotic Mackey-Glass time series forecasting at its predictability horizon, a task demanding long-range temporal memory and high noise robustness. Controlled, feedforward architectures were used to compare MTC, LIF, and SOTA AdLIF baselines under identical experimental and optimization conditions.

Key Quantitative Results

Predictive Fidelity:
- MTC achieves median $R^2 \approx 0.988$ and SMAPE $\approx 17\%$ on the forecasting task.
- AdLIF attains $R^2 \approx 0.982$ and SMAPE $\approx 21\%$ .
- LIF baseline only reaches $R^2 \approx 0.66$ , SMAPE $\approx 78\%$ .
Sparsity and Efficiency:
- MTC achieves an order of magnitude reduction in mean spike communication probability ( $S_{comm} \approx 2 \times 10^{-3}$ ), compared to AdLIF ( $S_{comm} \approx 4 \times 10^{-2}$ ).
- Computational inactivity ( $S_{comp} \gt 0.9$ for MTC) substantially exceeds AdLIF ( $S_{comp} \approx 0.7$ ) and LIF.

The MTC model therefore achieves nearly optimal forecasting fidelity while operating at the highest computational and communication sparsity among all tested SNN variants.

Encoding Strategies

Analysis demonstrates that AdLIF improves accuracy by increasing firing density, relying on regular, tonic spiking. In contrast, the MTC architecture utilizes its firing heterogeneity and multi-timescale integration to encode information with brief, temporally precise spike bursts and distributed codes—thereby compressing information and reducing activity without a significant loss in accuracy.

Importantly, the performance-efficiency tradeoff is not linear: MTC decisively outperforms both LIF and AdLIF, simultaneously advancing accuracy and efficiency. The data imply that conductance-based multi-timescale mechanisms fundamentally improve the quality of temporal representation and energy-awareness in SNNs.

Practical and Theoretical Implications

The MTC approach demonstrates the feasibility and advantage of gradient-trainable, multi-timescale conductance mechanisms for SNN-based temporal processing:

Hardware Integration: Conductance-based elements are suitable for analog neuromorphic realizations using standard transconductance blocks, supporting low-power, event-driven computation.
Sparsity-Driven Computation: Dynamic sparsity emerges naturally from conductance shaping instead of being enforced through external regularization or network constraints. This aligns with energy minimization goals for edge AI and embedded neuromorphic devices.
Temporal Coding: The increased firing regime diversity provides a more expressive single-neuron "vocabulary" for temporal pattern representation, which may support superior learning and robustness in sequential and control tasks.
Biological Plausibility: The mechanistic alignment with biophysical processes strengthens the theoretical foundation for deploying SNNs in domains that require robust, adaptive memory and noise resilience.

Limitations and Future Directions

Several salient issues are noted for future investigation:

The observed efficiency-accuracy gains are most pronounced in the tested regression setting; in event-driven spiking benchmarks, the tradeoffs between adaptation and conductance-based mechanisms may be task-dependent.
Further exploration of recurrent architectures, additional datasets, and spiking/non-spiking input coding schemes will be critical to fully characterize MTC performance envelopes.
Hardware proxy metrics (activity sparsity, communication events) must ultimately be validated against actual energy measurements in practical neuromorphic VLSI implementations.
Parameter learning for conductance timescales and systematic architectural initialization could yield further optimizations.
Rigorous dynamical analyses and long-horizon closed-loop evaluations are necessary to confirm long-term predictability and robustness.

Conclusion

The paper establishes multi-timescale conductance spiking neurons as a promising direction for energy-efficient, temporally expressive SNNs in continuous regression and dynamic temporal tasks. By conferring a continuum of biophysically grounded firing behaviors within a differentiable, trainable framework, MTC neurons achieve high accuracy and unmatched sparse computation. Conductance-based neuronal excitability thus presents a compelling alternative to conventional adaptation mechanisms in SNN design, with far-reaching implications for neuromorphic computing and temporal machine learning domains.