Dynamic Decay Spiking Neurons

• Dynamic decay spiking neurons are computational models that replace fixed decay with tunable, state-dependent decay mechanisms for improved integrative and threshold dynamics.
• They apply adaptive decay to membrane potentials, synaptic traces, and firing thresholds, enhancing biological plausibility and efficiency.
• Leveraging learnable decay parameters, these models achieve robust performance in image classification, robotics, and long-memory signal encoding tasks.

A dynamic decay spiking neuron is a computational model in which the subthreshold and/or threshold dynamics governing spike emission are modulated via tunable, state- or input-dependent decay mechanisms rather than fixed exponential decay. This family of models generalizes conventional integrate-and-fire neurons by introducing adaptable or learned decay processes for postsynaptic potentials, membrane potential integration, or firing threshold, often to increase expressivity, biological plausibility, homeostatic stability, or computational efficiency in spiking neural networks.

1. Mathematical Foundations of Dynamic Decay in Spiking Neurons

Dynamic decay spiking mechanisms regulate the time evolution of core neuronal state variables through non-constant decay laws. The classical leaky integrate-and-fire (LIF) neuron employs a static exponential decay for the membrane potential:

$$U[t] = \alpha U[t-1] + I[t] - V_\mathrm{th} S[t-1], \quad \alpha = \exp(-\Delta t/\tau)$$

where $\tau$ is the membrane time constant, $U[t]$ is the potential, $I[t]$ the input, and $S[t-1]$ the prior spike. Dynamic decay variants replace the fixed $\alpha$ with a state-dependent, learnable, or otherwise adaptive function. Examples include:

• Learnable polynomial decay: $f_\theta(U)$, parameterized as a polynomial, replaces the linear decay (Jahns et al., 7 Oct 2025).
• Dual decay (temporal and spatial): Separate decay coefficients for temporal retention and synaptic integration (Zhang et al., 5 Feb 2025).
• Cascaded exponentials for long-memory: Approximations of power-law decay to capture fractional, non-Markovian temporal dynamics (Bohte et al., 2010).
• Threshold decay: Adaptive threshold governed by energy or temporal change rates (Ding et al., 2022).

This flexibility allows the neuron to model rich subthreshold and suprathreshold behavior beyond the reach of fixed-decay LIF or SRM models.

2. Core Mechanisms and Model Instantiations

Dynamic decay appears in several rigorously defined models, each encoding decay in a specific neural subsystem:

Model/Framework	Dynamic Decay Location	Mechanism
FC (Firing Cell) (Bialowas et al., 2017)	Postsynaptic potential registers $x_k$ and synaptic calcium $C_k$	Exponential-like decay of PSP and LTP traces; discrete step
LNM (Jahns et al., 7 Oct 2025)	Membrane potential $U[t]$	Polynomial $f_\theta$-param decay (learnable by gradient)
DA-LIF (Zhang et al., 5 Feb 2025)	Temporal $\beta$, spatial $\alpha$ factors	Independently learnable tanh-param decays per layer
Fractional Spiking (Bohte et al., 2010)	Refractory/post-spike tail	Power-law decay, optionally via exponential cascades
BDETT (Ding et al., 2022)	Dynamic firing threshold	Energywise, temporally decaying threshold

Detailed mechanisms:

Firing Cell—Postsynaptic Dynamic Decay

Each synapse $k$ maintains a PSP register $x_k[n]$ obeying:

$x_k[n+1] = \gamma x_k[n] + A_k s_k[n]$

$\gamma = \exp(-\Delta t/\tau_{\mathrm{psp}})$, $A_k$ is spike amplitude. Each synapse also tracks a dynamic calcium trace $C_k[n]$, updated and decayed separately, driving synaptic weight change via long-term potentiation (LTP) (Bialowas et al., 2017).

Learnable Neuron Models (LNM)

Decay is expressed as a flexible, data-driven polynomial transformation:

$$U[t] = f_\theta\left(U[t-1]\right) + I[t] - V_{\mathrm{th}} S[t-1]$$

where

$$f_\theta(u) = \sum_{i=0}^N \theta_i \, u^i,\quad u\in[-1,1]$$

Parameters $\theta$ are trained with surrogate-gradient backpropagation (Jahns et al., 7 Oct 2025).

Dual Adaptive LIF (DA-LIF)

DA-LIF introduces two separate decays: $V^{t,n} = \beta^n H^{t-1,n} + \alpha^n X^{t,n}$ $\alpha^n$ controls spatial (synaptic integration) decay, $\beta^n$ controls temporal (membrane retention) decay, both parameterized via $\tanh(\phi)$ and learned per layer via STBP (Zhang et al., 5 Feb 2025).

Fractionally Predictive Neurons

The decay of the postsynaptic current or the reconstruction kernel follows a power-law: $\kappa_\alpha(t) = t^{-\alpha}/\Gamma(1-\alpha)$ allowing the neuron to approximate fractional derivatives and capture long-range temporal dependencies more efficiently than exponentials. Biological approximation uses cascades of exponentials (Bohte et al., 2010).

Dynamic Energy-Temporal Threshold (BDETT)

The neuron maintains a dynamic threshold

$$\Theta^{l}_i(t) = \frac{1}{2} \left[ E^l_i(t) + T^l_i(t) \right]$$

where $E^l_i(t)$ and $T^l_i(t)$ are energy- and temporally-based terms with state-dependent decay, calibrated to ensure homeostatic firing and adaptive responsiveness (Ding et al., 2022).

3. Implications for Network Computation, Homeostasis, and Plasticity

Dynamic decay mechanisms confer several computational advantages and network-level effects:

• Temporal Filtering and Adaptivity: Power-law or learnable decay endows neurons with finely tunable temporal integration and adaptive receptive fields. Slow decay enables long-term accumulation; rapid decay enforces sparse, temporally precise firing (Bohte et al., 2010, Jahns et al., 7 Oct 2025).
• Homeostatic Regulation: Dynamic decay—especially in threshold (BDETT)—maintains stable firing rates and effective homeostasis in the face of input variability, noise, or weight drift, outperforming static-threshold and heuristic baselines in real-world robotics and RL tasks (Ding et al., 2022).
• Plasticity and Memory: Coupling dynamic decay with short- and long-term synaptic traces (e.g., $C_k[n]$ in FC) yields models that support both STP (short-term potentiation) and LTP, directly relating decay rates to memory time scales and learning rates (Bialowas et al., 2017).
• Expressivity and Performance: Learnable decays (polynomial in LNM, dual in DA-LIF) improve accuracy in static and neuromorphic benchmarks, with demonstrable gains over fixed-leak models (e.g., $+1.19\%$ on CIFAR-100, $+2.01\%$ on ImageNet in LNM) with minimal parameter overhead (Jahns et al., 7 Oct 2025, Zhang et al., 5 Feb 2025).

4. Training, Implementation, and Hardware Realization

Modern dynamic decay neurons are fully trainable in deep spiking networks:

• Parameter Learning: Polynomial and per-layer decay coefficients are trained via surrogate gradients, chain rule through the dynamic decay function, and standard optimizers (SGD+momentum). Additional regularization (weight decay, clipping) is needed for stability (Jahns et al., 7 Oct 2025, Zhang et al., 5 Feb 2025).
• Initialization and Constraints: Initialization to identity or standard LIF decay, enforcement of $f_\theta(0)=0$, and input clipping ensure numerical robustness (Jahns et al., 7 Oct 2025).
• Complexity and Overhead: DA-LIF and LNM add only $O(L)$ parameters (number of layers) and ~3–5% compute/energy overhead compared to fixed-decay LIF, while maintaining SNN efficiency (Zhang et al., 5 Feb 2025).
• Neuromorphic Suitability: The use of shift-registers (FC), exponential cascades, or local dynamic variables (BDETT) are well-suited to efficient hardware (analog or digital neuromorphic, event-driven architectures) due to locality, simplicity, and batch/statistics-driven operations (Bialowas et al., 2017, Ding et al., 2022, Bohte et al., 2010).

5. Experimental Validation and Comparative Performance

Dynamic decay spiking neurons have been empirically validated in a range of tasks:

• Image Classification: DA-LIF and LNM consistently outperform fixed-decay SNNs. Example: DA-LIF achieves $96.72\%$ ($T=4$) on CIFAR-10 (Zhang et al., 5 Feb 2025); LNM achieves $97.01\%$ ($T=4$) (Jahns et al., 7 Oct 2025).
• Robust Control and Robotics: BDETT shows $10-12\%$ absolute improvement in robot obstacle avoidance and large reduction in firing-rate variance under degraded input/quantization (Ding et al., 2022).
• Long-Memory Signal Encoding: Fractionally predictive neurons require half the spikes of single-exponential models for comparable SNR on fractal signals, leveraging power-law kernel dynamics (Bohte et al., 2010).
• Ablation Studies: Both DA-LIF and LNM report that higher-order polynomial or dual decays yield significant incremental accuracy, confirming the utility of nontrivial decay parameterizations (Jahns et al., 7 Oct 2025, Zhang et al., 5 Feb 2025).

6. Biological Plausibility and Connections

Dynamic decay paradigms align with numerous empirical and theoretical observations:

• Biological Heterogeneity: Dual-adaptive and learnable decay mechanisms directly map to observed diversity in time and spatial integration across cortical neuron subtypes (Zhang et al., 5 Feb 2025).
• Adaptive Thresholds: The BDETT model leverages mechanisms inferred from barn-owl IC and mammalian cortex, where threshold is modulated by both mean depolarization and rapid voltage changes (Ding et al., 2022).
• Power-law Adaptation: Fractional (power-law) decay matches empirically measured adaptation exponents ($\beta\approx0.8$), and supports a unified view of firing as predictive fractional differentiation (Bohte et al., 2010).

A plausible implication is that dynamic decay architectures close the gap between theoretical SNNs and the complexity of biophysical neuron dynamics, supporting their use as computational substrates for both brain-like and robust, energy-efficient artificial systems.

7. Outlook and Integration Strategies

The deployment of dynamic decay spiking neurons in advanced SNNs rests on:

• Drop-in Replacement: Models like BDETT and DA-LIF can replace fixed-decay neurons in existing frameworks using straightforward updates and minimal hyperparameter tuning (Ding et al., 2022, Zhang et al., 5 Feb 2025).
• Tuning: Decay parameters can be optimized via grid-search or learned end-to-end. Key choices include polynomial degree ($N$ in LNM), number of cascaded exponentials (fractional models), and initialization range.
• Layerwise/Neuronwise Customization: Per-layer or even per-neuron decay parameterization supports heterogeneous processing, matching both biological and computational requirements (Zhang et al., 5 Feb 2025).

Dynamic decay spiking neurons thus constitute a foundational component in next-generation SNN research and deployment, supporting high expressivity, homeostasis, and efficient temporal computation across a wide array of platforms and tasks.