Memory-Augmented Spiking Networks: Synergistic Integration of Complementary Mechanisms for Neuromorphic Vision

Effiong Blessing; Chiung-Yi Tseng; Isaac Nkrumah; Junaid Rehman

cs.NEFeb 21, 2026

Memory-Augmented Spiking Networks: Synergistic Integration of Complementary Mechanisms for Neuromorphic Vision

Authors: Effiong Blessing, Chiung-Yi Tseng, Isaac Nkrumah, Junaid Rehman

Organizations: Department Of Computer Science, Saint Louis University · Luxmuse AI · Saint Louis University · Independent Researcher

Abstract

Spiking Neural Networks (SNNs) provide biological plausibility and energy efficiency, yet systematic investigations of memory augmentation strategies remain limited. We conduct a five-model ablation study integrating Leaky Integrate-and-Fire neurons, Supervised Contrastive Learning (SCL), Hopfield networks, and Hierarchical Gated Recurrent Networks (HGRN) on the N-MNIST dataset. Baseline SNNs exhibit organized neuronal groupings, or structured assemblies, characterized by a silhouette score of $0.687 \pm 0.012$ . Individual augmentations introduce trade-offs: SCL improves accuracy by $0.28\%$ but reduces clustering (silhouette score $0.637 \pm 0.015$ ), while HGRN yields consistent gains in both accuracy ( $+1.01\%$ ) and computational efficiency ( $170.6\times$ ). Full integration achieves a balanced improvement across metrics, reaching a silhouette score of $0.715 \pm 0.008$ , classification accuracy of $97.49 \pm 0.10\%$ , energy consumption of $1.85 \pm 0.06\,μ\mathrm{J}$ , and sparsity of $97.0\%$ . These results indicate that optimal performance emerges from architectural balance rather than isolated optimization, establishing design principles for memory-augmented neuromorphic systems.

Figures & tables

Fig. 1 : Complete Memory-Augmented SNN Architecture. Three-stage pipeline integrating: Stage 1 (SNN Encoder, blue) processes dynamic vision sensor(DVS) spikes through Conv1+LIF, Conv2+LIF, and FC(512)+LIF feature extraction; Stage 2 (Memory Augmentation, green) applies Supervised Contrastive Learning for engram formation and Hopfield networks for associative memory; Stage 3 (HGRN Temporal Gating, red) performs selective information flow with forget/update/cell gates over T=25 steps.

Fig. 2 : LIF Neuron Temporal Dynamics and Circuit Model. Top: Membrane potential

U(t)

exhibits leaky integration with decay

\beta=0.9

, generating spikes when crossing threshold

\theta=1.0

. Bottom: Circuit model showing synaptic inputs converging at soma with membrane capacitance and leak, implementing biological integrate-and-fire dynamics.

Fig. 3 : Hopfield Network Module for Associative Memory. Top flowchart: Six-stage process from Input Pattern (blue, $h^{0}$ from SNN, 512-D with dot pattern visualization) through Weight Matrix (green, $\mathbf{W}=\sum\boldsymbol{\xi}^{p}(\boldsymbol{\xi}^{p})^{T}$ , outer-product learning), Energy (red, $E=-\frac{1}{2}\boldsymbol{h}^{T}\mathbf{W}\boldsymbol{h}$ , minimization to attractors), State Update (purple, $h^{t+1}=\text{sign}(\mathbf{W}h^{t})$ , recurrent iterations t=0,1,2), Convergence (yellow, check $||h^{t+1}-h^{t}||<\epsilon$ , max $T_{max}=5$ ), to Retrieved Pattern (blue, $h^{*}$ , 512-D). Purple dashed arrow shows recurrent loop. Bottom visualizations: Left shows energy landscape with global attractor (red star), local minima (blue triangles), and convergence paths (purple arrows) in 2D state space; Center shows stored pattern assemblies for digits 0-9 as colored clusters with overlap enabling generalization; Right shows convergence dynamics with pattern similarity (blue, 0 $\rightarrow$ 1.0) and energy (red, 12 $\rightarrow$ 1.2) over 5 iterations, demonstrating rapid convergence and energy minimization.

Fig. 4 : Information Flow Control Mechanism in HGRN. Vertical flowchart showing how Previous State $C_{t-1}$ (green) influences current processing through two pathways: Forget Gate $f_{t}$ (orange, left) determines what to selectively discard, and Update Gate $u_{t}$ (orange, right) determines what to selectively integrate. Both gates converge to Current State $c_{t}$ (yellow) through element-wise operations. Formula box (pink) shows the cell state update equation $c_{t}=f_{t}\odot c_{t-1}+u_{t}\odot\tilde{c}_{t}$ combining forget and update operations. Arrow flows to Next State $c_{t+1}$ (purple), completing the temporal recurrence.

Fig. 5 : HGRN Temporal Gating Mechanism - Detailed Module. Top: Information flow from Memory State (green, 512-D from Hopfield) through Forget Gate

f_{t}

(pink, controls what to discard) and Update Gate

u_{t}

(pink, controls what to integrate) to Cell State Update (yellow, gated information flow), Hidden State (blue, goal-directed representation), Hierarchical Coordination (gold, multi-level context integration), and Output (purple, to classifier). Purple dashed line shows temporal recurrence

h_{t-1}\rightarrow h_{t}

maintaining memory across time. Bottom panels: Left shows complete gate equations with sigmoid and element-wise operations; Center shows HGRN performance benefits including +1.01% accuracy improvement, 0.698 silhouette, 1.85 µJ energy (170.6

\times

vs ANNs), 97.0% sparsity, differentiable gradients, and temporal alignment with spike dynamics; Right explains why HGRN works through temporal domain operation, smooth gradients (unlike Hopfield sign function), adaptive temporal filtering, and consistent improvements across all metrics.

Model	Val Acc (%)	Test Acc (%)	Silhouette	Energy (µJ)
M1: Baseline (No SCL)	96.43	96.35	0.687	2.39
M2: Baseline + SCL	96.71	96.68	0.637	2.90
M3: SNN + SCL + Hopfield	96.21	96.15	0.695	$\sim$ 2.3
M4: SNN + SCL + HGRN	97.44	97.38	0.698	1.85
M5: Full Hybrid (All)	97.49	97.44	0.715	$\sim$ 1.9

TABLE I : Comprehensive Ablation Study Results on N-MNIST. Test set accuracies reported. Best validation accuracies: M1 (96.67%), M2 (96.94%), M3 (96.85%), M4 (97.57%), M5 (97.54%).

Fig. 6 : Training Dynamics Across Five Configurations. Validation accuracy (top-left) shows Models 4 and 5 achieving the highest performance with stable convergence. Training loss (top-right) demonstrates smooth optimization. Final validation accuracies: M1 (96.67%), M2 (96.94%), M3 (96.85%), M4 (97.57%), M5 (97.54%).

Fig. 7 : Memory Assembly Quality Comparison Across Configurations. Silhouette scores reveal unexpected component interactions: Baseline (M1) achieves good clustering (0.687) through spike-timing dynamics alone. SCL (M2) disrupts structure (0.637) despite improving accuracy. Memory mechanisms (M3/M4) partially restore quality (0.695-0.698). Full integration (M5) achieves excellent clustering (0.715, marked with star), exceeding the 0.7 “excellent” threshold and demonstrating synergistic effects through architectural balance.

Panel of Figure 9 — Fig. 8 : t-SNE Visualization of Memory Assembly Evolution. (a) Baseline (M1, Silhouette 0.687): SNNs naturally form well-separated class clusters through spike-timing dynamics, showing distinct but somewhat overlapping assemblies. (b) +SCL (M2, Silhouette 0.637): Contrastive learning disrupts existing structure, increasing cluster overlap as output-space optimization conflicts with spike-timing organization. (c) Full Hybrid (M5, Silhouette 0.715): Synergistic integration achieves tightly grouped, excellently separated clusters, demonstrating that architectural balance resolves individual component conflicts.(See M3, M4 in appendix)

Model	Energy/Inf	Total Energy	Sparsity	Reduction	Energy Saved	SynOps	MACs
	(µJ)	(µJ)	(%)	vs ANN	(%)	(M)	(M)
M1: Baseline	0.0030	4.72	98.73	403.1 $\times$	99.8	5.247	413.84
M2: +SCL	0.0036	5.79	98.44	328.6 $\times$	99.7	6.438	413.84
M3: +Hopfield	0.0039	6.30	98.31	302.0	99.7	7.004	413.84
M4: +HGRN	0.0020	3.15	99.15	603.9 $\times$	99.8	3.503	413.86
M5: Full Hybrid	$\sim$ 0.0020	$\sim$ 3.2	97.0	$\sim$ 600 $\times$	99.8	$\sim$ 3.5	$\sim$ 414

TABLE II : Comprehensive Energy Efficiency Analysis

Method	Accuracy (%)	Bio-plausible
Diehl & Cook (2015) [ 3 ]	$\sim$ 95.0	Yes (STDP)
SLAYER (2018) [ 12 ]	92.5	Yes
Wu et al. (2018) [ 13 ]	99.4	Partial
ANN-SNN (2021) [ 14 ]	98.5	Partial
Ours (M4: SNN+HGRN)	97.44	Yes
Ours (M5: Full Hybrid)	97.49	Yes

TABLE III : Comparison with State-of-the-Art Methods on N-MNIST

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 12 — Fig. 9 : Cluster Analysis for Models 3 and 4. (a) Hopfield integration partially restores structure (0.695 vs M2’s 0.637). (b) HGRN achieves better clustering (0.698), demonstrating temporal gating integrates naturally with spike processing. It shows detailed t-SNE visualizations for Models 3 and 4, complementing the main paper’s analysis of Models 1, 2, and 5.

Spiking Neural Networks

Spiking neural networks process information through discrete, time-dependent neural events, aiming for efficient computation and biologically inspired learning.

See all 351 papers

Saliency-Guided Replay for Continual Learning in Deep Spiking Neural Networks
Oct 13, 2026Francesco Panetta, Federico Di Valerio, Michela ProiettiStorage-Efficient Continual Learning Replay-Based Continual Learning
Sapienza University of Rome, Italy · Goethe University Frankfurt, Germany
Noisy group neurons with synchronous resetting for high-performance spiking neural networks
Aug 18, 2026Yajie Zhai, Yanmei Kang, Meng Li +1Spiking Neural Networks Neuromorphic Computing
Department of Applied Mathematics, School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an 710049, China · Center for Intersection of Mathematics and Life Sciences, Xi’an Jiaotong University, Xi’an 710049, China · The Key Laboratory of Biomedical Information Engineering of Ministry of Education, Institute of Health and Rehabilitation Science, School of Life Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China +1
Efficient Memristive Spiking Neural Networks Architecture with Supervised In-Situ STDP Method
Jul 28, 2025Santlal Prajapati, Susmita Sur-Kolay, Soumyadeep DuttaEfficient Neural Network Inference Spiking Neural Networks
Advanced Computing and Microelectronics Unit, Indian Statistical Institute, Kolkata 700108, India · Balasore Alloys Limited, Balgopalpur, Odisha 756020, India

Oct 13, 2026J/K move · Enter open · S save

Saliency-Guided Replay for Continual Learning in Deep Spiking Neural Networks

Francesco Panetta, Federico Di Valerio, Michela Proietti

Sapienza University of Rome, Italy · Goethe University Frankfurt, Germany

Storage-Efficient Continual Learning Replay-Based Continual Learning Spiking Neural Networks Incremental Learning

Open paper AskPDF arXiv

Spiking Neural Networks (SNNs) are promising for energy-efficient neuromorphic computing, but they still suffer from catastrophic forgetting when learning tasks sequentially. This work investigates the integration of Experience Packing and Replay (EPR) as a memory-efficient continual learning strategy for deep SNNs under low-latency constraints. The proposed framework combines Batch Normalization Through Time stabilized VGG9 SNNs with Spike Activation Maps (SAM), which provide forward-only explanations based on spike activity. These explanations are temporally aggregated into spatial attention maps and used to select the most informative image patches for replay, instead of storing full images. Experiments on Split CIFAR10 and Split CIFAR100 show that SAM-guided EPR improves the accuracy-memory trade-off under tight memory budgets, and remains competitive while requiring substantially less storage.

Abstract

Spiking Neural Networks

Explore similar work

Saliency-Guided Replay for Continual Learning in Deep Spiking Neural Networks

Noisy group neurons with synchronous resetting for high-performance spiking neural networks

Efficient Memristive Spiking Neural Networks Architecture with Supervised In-Situ STDP Method