Learning While Inferring: Local and Parallel Learning for Edge SNNs across Sensing Modalities

Yanxun Zhang; Yifei Wang; Changze Lv; Jingwen Xu; Yiyang Lu; Xiaohua Wang; Di Yu; Xin Du

cs.NEOct 2, 2026

Learning While Inferring: Local and Parallel Learning for Edge SNNs across Sensing Modalities

Authors: Yanxun Zhang, Yifei Wang, Changze Lv, Jingwen Xu, Yiyang Lu, Xiaohua Wang, Di Yu, Xin Du,

Organizations: College of Computer Science and Artificial Intelligence, Fudan University · Shanghai Key Laboratory of Intelligent Information Processing · School of Software Technology, Zhejiang University

Abstract

Edge intelligence requires models to sense continuously in real time and to keep adapting on-device, all under tight compute, energy, and memory budgets. Although spiking neural networks (SNNs) enable efficient event-driven inference, standard surrogate-gradient backpropagation (BP) serializes updates and blocks ongoing inference. We investigate Bidirectional Spike-Based Distillation (BSD) as an on-device learning principle that lets edge SNNs learn while inferring. BSD couples a stimulus-driven forward network with an independent target-driven reverse network and aligns their intermediate representations through local objectives. Because the two pathways have disjoint computation graphs until alignment, forward inference, reverse inference, and stage-wise updates can run concurrently. On 25 benchmarks spanning the five sensing modalities of SOUL, BSD stays within 3.8 percentage points of matched BP baselines on average, while only its forward branch is needed at deployment. The learned representations also transfer well to few-shot class-incremental learning without replaying past data. By removing the dense floating-point backward chain, BSD reduces the projected training latency to $0.72\times$ and the estimated training energy to $0.36\times$ that of BP.

Figures & tables

Figure 1: Overview of edge learning with BSD. Left: heterogeneous edge sensing modalities. (a) BSD couples a stimulus-driven forward (inference) branch with a target-driven reverse (training) branch aligned by local ReCo objectives. (b) The training pipeline of BSD and BP. BSD executes inference, reverse inference, and local updates concurrently, unlike the strictly sequential BP pipeline below. (c) BSD achieves accuracy comparable to BP across the five modalities while requiring substantially less training energy and a shorter projected training time.

Dataset	BP		BSD (Ours)
Vision
MNIST	99.56	$\pm$ 0.14	99.44	$\pm$ 0.03
Fashion-MNIST	92.68	$\pm$ 0.42	91.05	$\pm$ 0.20
CIFAR-10	87.12	$\pm$ 1.76	84.13	$\pm$ 0.34
CIFAR-100	57.75	$\pm$ 0.35	53.48	$\pm$ 0.22
SVHN	94.36	$\pm$ 0.73	90.81	$\pm$ 0.11

Table 1: Test accuracy (%) of BP and BSD with matched forward architectures (mean

\pm

std, three runs). Models are convolutional SNNs unless marked R (spiking RNN) or L (spiking LSTM).

Figure 2: One CIFAR-10 training step drawn to scale (Table 5 ). BP runs its phases in sequence and blocks inference; BSD overlaps its branches and splits local work over

P=6

workers.

Figure 3: BSD/BP training-cost ratios on ten representative benchmarks (two per modality): operation-level energy (left) and projected time under concurrent execution (right); values below 1 mean that BSD is cheaper.

Figure 4: FSCIL on CIFAR-100: seen accuracy (left) and H score (right) per session, mean

\pm

std over three runs.

Variant	Seen Acc.		H Score
Full pipeline	42.18	$\pm$ 0.60	32.54	$\pm$ 0.36
w/o prototype TTA	40.33	$\pm$ 0.31	29.14	$\pm$ 0.51
w/o multi-layer fusion	38.47	$\pm$ 0.12	31.86	$\pm$ 0.28
w/o subspace proj.	41.63	$\pm$ 0.45	32.26	$\pm$ 0.29

Table 2: Ablation of the adaptation components in the final FSCIL session (%, mean

\pm

std over three runs).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Temporal BSD (Section 4.2 ) unrolled over time. The forward RNN runs from

t=1

T

on the inputs

x_{t}

, the reverse RNN runs from

t=T

1

on the constant target

e_{y}

, and ReCo aligns their hidden states

h_{t}

and

q_{t}

at every step; the task loss acts only on the time-averaged forward readout. Detaching the recurrent states every

n

steps limits temporal credit assignment to segments of length

n

Modality	Dataset	Model	Batch size	Learning rate	Weight decay	$U_{\mathrm{thr}}^{F}$ / $U_{\mathrm{thr}}^{B}$
Vision	MNIST †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8
	Fashion-MNIST †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8
	CIFAR-10 †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8
	CIFAR-100 †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8
	SVHN †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8
	Tiny-ImageNet †	CNN	128	$10^{-4}$	$10^{-5}$	0.8/0.8

Appendix

Table 3: Per-dataset training configurations.

U_{\mathrm{thr}}^{F}

and

U_{\mathrm{thr}}^{B}

are the firing thresholds of the forward and reverse branches. Rows marked † follow the configurations of the original BSD paper ( Wang et al., 2026 ) .

Figure 6: Prototype-based FSCIL on the frozen BSD forward branch (Section 4.3 ). In each incremental session, novel prototypes are built from view-averaged support features at two layers and projected toward the subspace of the base prototypes; a query is classified by its fused cosine scores against all prototypes. Numbered markers indicate the components ablated in Table 2 .

		MACs ( $\times 10^{6}$ )		ACs ( $\times 10^{6}$ )		Energy (mJ)		Ratio
Modality	Dataset	BP	BSD	BP	BSD	BP	BSD	BSD/BP
Vision	MNIST	106	27.2	45.9	63.8	0.528	0.182	0.345
	Fashion-MNIST	122	27.2	53.9	76.1	0.609	0.194	0.318
	CIFAR-10	158	36.8	57.9	78.2	0.780	0.240	0.307
	CIFAR-100	157	36.8	57.4	76.0	0.775	0.238	0.307
	SVHN	150	36.8	53.8	70.6	0.738	0.233	0.315

Appendix

Table 4: Per-sample operation counts and estimated training energy of the matched BP and BSD models. MACs and ACs are given in millions and energy in mJ, all with three significant figures.

		$T_{F}$		$T_{R}$	$T_{L}$		$T_{U}$		$T_{\mathrm{opt}}$		Ratio
Modality	Dataset	BP	BSD	BSD	BP	BSD	BP	BSD	BP	BSD	BSD/BP
Vision	MNIST	16.0	21.6	21.9	17.7	62.6	2.08	2.63	35.8	32.8	0.915
	Fashion-MNIST	18.8	14.8	20.6	18.8	64.0	2.09	2.48	39.7	31.7	0.798
	CIFAR-10	14.9	12.5	14.1	21.2	71.2	1.76	2.42	37.8	26.4	0.697
	CIFAR-100	16.4	16.1	17.8	23.7	77.8	1.84	3.06	41.9	31.3	0.747
	SVHN	12.4	12.9	14.8	17.3	99.4	2.01	2.68	31.7	31.8	1.004

Appendix

Table 5: Profiled phase times per training step (ms) and the resulting step times,

T_{\mathrm{opt}}(\mathrm{BP})=T_{F}+T_{L}+T_{U}

and

T_{\mathrm{opt}}(\mathrm{BSD})=\max(T_{F},T_{R})+(T_{L}+T_{U})/P

T_{R}

applies only to BSD, since BP has no reverse branch.

		Training batch size
Modality	Dataset	64	32	16	8	4
Vision	CIFAR-10	81.5	78.1	69.0	50.0	36.1
Motion	MotionSense	95.7	95.1	94.5	90.5	79.4
Acoustic	GTZAN	61.5	60.5	60.0	51.0	43.5
Wireless	AOPHand	80.6	76.8	75.8	71.2	60.1
Neuromorphic	SHD	98.5	96.5	93.9	84.9	83.0

Appendix

Table 6: BSD test accuracy (%) under different training batch sizes. Performance is largely preserved down to a batch size of

32

and degrades noticeably below

16

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

Spiker-LL: An Energy-Efficient FPGA Accelerator Enabling Adaptive Local Learning in Spiking Neural Networks
May 18, 2026Alessio Caviglia, Filippo Marostica, Alessandro Savino +1On-Device Training Energy-Efficient ML
Control and Computer Engineering Department, Politecnico di Torino, Turin, Italy
L-SPINE: A Low-Precision SIMD Spiking Neural Compute Engine for Resource-efficient Edge Inference
Apr 4, 2026Sonu Kumar, Mukul Lokhande, Santosh Kumar VishvakarmaEfficient Neural Network Inference Spiking Neural Networks
Centre for Advanced Electronics, Indian Institute of Technology Indore, Simrol 453552, Indore, India · Dept. of Electrical Engineering, Indian Institute of Technology Indore, Simrol 453552, Indore, India
EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures
Apr 29, 2026Gustav Olaf Yunus Laitinen-Fredriksson Lundstrom-Imanov, Taner YilmazEnergy-Efficient ML Edge Inference
Department of Economics, Stockholm University, Universitetsvägen 10 A, SE-106 91 Stockholm, Sweden · Department of Computer Engineering, Afyon Kocatepe University, 03200 Afyonkarahisar, Türkiye

May 18, 2026J/K move · Enter open · S save

Spiker-LL: An Energy-Efficient FPGA Accelerator Enabling Adaptive Local Learning in Spiking Neural Networks

Alessio Caviglia, Filippo Marostica, Alessandro Savino, Stefano Di Carlo

Control and Computer Engineering Department, Politecnico di Torino, Turin, Italy

On-Device Training Energy-Efficient ML Spiking Neural Networks Edge ML FPGA-Based Neural Network Acceleration On-Device Inference

Open paper AskPDF arXiv

Deploying adaptive intelligence at the edge remains challenging due to the high computational and energy cost of training neural models. Spiking Neural Networks (SNNs) offer a promising alternative, but enabling on-device learning requires hardware-algorithm co-design. This paper presents SPIKER-LL, an FPGA-based SNN accelerator that extends the open-source Spiker+ inference architecture with efficient support for the STSF local learning rule. Through targeted microarchitectural extensions, SPIKER-LL performs inference and online learning with minimal overhead. Across MNIST, F-MNIST, and DIGITS, it achieves up to 93% accuracy, sub-millisecond latency, and less than 0.1 mJ per inference, while remaining DSP-free and highly scalable for edge-FPGA deployments.

Abstract

Spiking Neural Networks

Explore similar work

Spiker-LL: An Energy-Efficient FPGA Accelerator Enabling Adaptive Local Learning in Spiking Neural Networks

L-SPINE: A Low-Precision SIMD Spiking Neural Compute Engine for Resource-efficient Edge Inference

EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures