Spike Agreement-Dependent Plasticity (SADP)

• Spike Agreement-Dependent Plasticity (SADP) is a synaptic learning rule that computes global spike train agreement using metrics like Cohen’s kappa to update weights and delays.
• It achieves linear-time complexity by processing binned spike data, enabling efficient implementation and scalability in neuromorphic hardware.
• SADP supports both supervised and unsupervised learning, delivering superior performance on benchmarks such as MNIST compared to classical STDP and Hebbian rules.

Spike Agreement-Dependent Plasticity (SADP) defines a class of synaptic learning rules for spiking neural networks (SNNs) that drive synaptic updates based on the agreement between the full pre- and post-synaptic spike trains, as opposed to the precise timing of individual spike pairs. SADP generalizes classical spike-timing-dependent plasticity (STDP) to the population level, achieves linear-time complexity, and has demonstrated superior accuracy and hardware suitability in both unsupervised and supervised SNNs. This paradigm encompasses a family of mechanisms, including statistical-agreement-based weight plasticity and activity-dependent delay alignment, and admits both biological plausibility and scalability for large-scale neuromorphic systems (Bej et al., 22 Aug 2025, Farner et al., 2022, S et al., 13 Jan 2026, Tian et al., 8 Dec 2025).

1. Mathematical Foundations and Update Rules

SADP replaces the pairwise spike-timing dependency of traditional STDP with an aggregate measure of agreement between vectors of binned spikes over a time window $T$. For a synapse connecting pre-synaptic neuron $i$ and post-synaptic neuron $j$, denote their spike trains for batch sample $b$ as $\mathbf{x}_{b,i}, \mathbf{s}_{b,j} \in \{0,1\}^T$. The core update is driven by Cohen’s kappa coefficient:

$$p_0^{(b)} = \frac{1}{T} \sum_{t=1}^T \mathbf{1}\left(x_{b,i,t} = s_{b,j,t}\right)$$

$$p_e^{(b)} = \left(\frac{1}{T}\sum_t x_{b,i,t}\right)\left(\frac{1}{T}\sum_t s_{b,j,t}\right) + \left(\frac{1}{T}\sum_t (1-x_{b,i,t})\right)\left(\frac{1}{T}\sum_t (1-s_{b,j,t})\right)$$

$$\kappa_{ij}^{(b)} = \frac{p_0^{(b)} - p_e^{(b)}}{\max(1 - p_e^{(b)}, \epsilon)}$$

This agreement metric is passed through a bounded learning function $\mathcal{L}: [-1,1] \to \mathbb{R}$ (e.g., piecewise-linear or spline-based), then batched:

$$\Delta w_{ij}^{(t)} = \frac{\eta_t}{B} \sum_{b=1}^B \mathcal{L}(\kappa_{ij}^{(b)})$$

$$w_{ij}^{(t+1)} = \mathrm{clip}\left\{ \mathrm{sign}(w_{ij}^{(t)}+\Delta w_{ij}^{(t)})\max(|w_{ij}^{(t)}+\Delta w_{ij}^{(t)}|, \epsilon), -1, 1 \right\}$$

In the supervised SADP framework, the post-synaptic reference can be a teacher or correct-class spike train, allowing strict synaptic locality and fully local supervised learning without backpropagation (Bej et al., 22 Aug 2025, S et al., 13 Jan 2026).

Delay-based SADP (also called delay learning) operates by adjusting synaptic delays to cluster spike arrivals at each post-neuron tightly in time. For each causal pre-spike arriving just before a post-spike:

$$\Delta d_{i,j} = -\alpha\tanh\left(\frac{(t_i + d_{i,j}) - \bar{t}_{\mathrm{pre}}}{\tau}\right)$$

where $\bar{t}_{\mathrm{pre}}$ is the mean arrival time of all contributing pre-spikes within a causal window (Farner et al., 2022).

2. Relation to and Differentiation from Spike-Timing-Dependent Plasticity

STDP prescribes weight changes based on each possible pre–post spike pair via a kernel $K_{\mathrm{STDP}}(\Delta t)$, enforcing strict temporal causality and requiring $O(S^2)$ operations (with $S$ the number of spikes). By contrast, SADP:

• Computes updates based on global agreement (alignment or anti-alignment) across spike trains, not pairwise relations.
• Incurs only $O(T)$ complexity per synapse per batch ($T$ time bins), independently of spike count.
• Potentiates for high positive $\kappa$ (spike-train alignment), depresses for negative $\kappa$, with no sensitivity to precise ordering.

SADP thereby generalizes STDP to a population-based, robust paradigm that maintains local plasticity but scales efficiently and admits noisy or hardware-imprecise implementations (Bej et al., 22 Aug 2025, S et al., 13 Jan 2026, Tian et al., 8 Dec 2025).

3. Algorithmic Realization and Hardware Considerations

SADP is implementable using simple operations on bit-vectors and averages, rendering it well-suited to neuromorphic hardware:

• The core agreement measure $(x_t == s_t)$ per time bin is an XNOR + popcount, efficiently executed in parallel.
• Learning functions $\mathcal{L}$ can be implemented as LUTs (for linear/spline kernels) or small embedded splines for device-aligned dynamics.
• Empirical device conductance traces (from iontronic organic memtransistors) are directly fit by spline kernels $f_{+}, f_{-}$, enabling hardware-calibrated update rules.

Delay-adjustment SADP requires detecting pre–post alignments, updating delays, and optionally constraining them, again using strictly local operations (Bej et al., 22 Aug 2025, Farner et al., 2022).

Pseudocode templates for both weight-based and delay-based SADP have been established in the literature, with all updates strictly local and parallelizable. See Table 1 for computational cost comparison.

Rule	Operation Complexity	Hardware Suitability
Pairwise STDP	$O(S^2)$	Costly, precise timing
SADP (weight)	$O(T)$	Efficient, bitwise ops
SADP (delay)	$O(L)$ per spike	Simple, local only

4. Experimental Performance and Applications

Unsupervised and supervised SADP have been empirically evaluated on standard vision and biomedical benchmarks (Bej et al., 22 Aug 2025, S et al., 13 Jan 2026):

• Unsupervised SADP (400-feature layer) on MNIST achieved $91.06\%$ accuracy with ideal spline kernel ($\sim 720$–$760$ s/epoch), compared to $12.94\%$ for pairwise STDP ( $2611$ s/epoch) and $11.35\%$ Hebbian.
• Fashion-MNIST: $78.73\%$ (ideal spline), $10.69\%$ (STDP).
• Supervised SADP within hybrid CNN–SNNs (1SADP, $T=25$): MNIST $99.09\%$, FMNIST $89.69\%$, CIFAR-10 $70.61\%$ with rapid convergence ($<20–30$ epochs).
• Device-inspired spline kernels introduce only modest accuracy loss ($\sim 2$–$3\%$).

Delay-based SADP in layerwise SNNs with fixed weights, trained on downscaled MNIST, led to marked improvements in pattern separation and generalization—most untrained networks scored at chance; after training, the majority reached $60$%–$100$% accuracy on seen classes, with substantial reduction in totally non-separable outputs (Farner et al., 2022).

Synchrony- and agreement-gated variants (DA-SSDP) further enable batch-level, loss-modulated update strengths, providing regularization in deep SNNs with negligible computational overhead and small but consistent accuracy gains on CIFAR-10, CIFAR-100, CIFAR10-DVS, and ImageNet-1K (Tian et al., 8 Dec 2025).

5. Biological Plausibility and Theoretical Significance

SADP models move beyond the canonical pairwise-causality perspective of STDP:

• Agreement-based rules capture experimental findings that synchrony and co-activity, not only microsecond-level sequence order, drive synaptic plasticity.
• The use of Cohen’s kappa is consistent with multi-factor modification, where chance-corrected correlation, membrane voltage, or local population activity modulate plasticity.
• Delay adjustment (as in (Farner et al., 2022)) has independent biological support—activity-dependent axonal and dendritic delay plasticity is documented in vivo, and theoretical analyses (Izhikevich 2006) indicate a major potential for enhancing network polychronization.

A plausible implication is that SADP-based rules may underpin population-level learning and memory mechanisms in biological neural circuits, enabling both robust, memory-rich coding and efficient adaptation.

6. Extensions, Limitations, and Outlook

SADP rules admit various extensions and open directions:

• Supervised variants enable fully local error-guided learning, neither requiring backpropagation nor explicit teacher currents (S et al., 13 Jan 2026).
• Agreement measures other than Cohen’s kappa (e.g., alternative synchrony metrics, kernelized comparisons) may be substituted.
• DA-SSDP introduces three-factor gated SADP, embedding reward or global performance modulation via dopaminergic scaling (Tian et al., 8 Dec 2025).
• Hardware-oriented kernels derived from physical devices further demonstrate SADP’s compatibility with emerging neuromorphic substrates (Bej et al., 22 Aug 2025).

Limitations include the reliance on binned spike-train comparisons, potential sensitivity to coding schemes (e.g., rate vs. latency coding), and the necessity for further evaluation on more complex tasks and biological systems.

7. Comparative Summary

SADP reconciles competing desiderata for SNN learning: strict synaptic locality, population-level robustness, and computational and hardware scalability. Empirical results indicate clear advantages in both runtime and learning performance over pairwise STDP and Hebbian baselines on standard benchmarks. The framework encompasses both weight and delay plasticity, supports both unsupervised and supervised regimes, and is theoretically and experimentally grounded for neuromorphic implementation (Bej et al., 22 Aug 2025, Farner et al., 2022, S et al., 13 Jan 2026, Tian et al., 8 Dec 2025).