FF-STDP: A Forward-Forward STDP-Based Energy-Efficient SNN Accelerator with On-Chip Learning

XU Changqing; MO Guiqi; LIU Yi; LI Yuke; CHEN Zixin; YANG Yintang

doi:10.12263/DZXB.20260284

您当前的位置：

首页 >

文章列表页 >

FF-STDP: A Forward-Forward STDP-Based Energy-Efficient SNN Accelerator with On-Chip Learning

Innovative Architectures and Electronic Systems for Efficient Artificial Intelligence | 更新时间：2026-10-16

- FF-STDP: A Forward-Forward STDP-Based Energy-Efficient SNN Accelerator with On-Chip Learning
- ACTA ELECTRONICA SINICA Vol. 54, Issue 6, Pages: 2443-2453(2026)
- 作者机构：
  
  1.西安电子科技大学广州研究院，广东广州 510555
  2.西安电子科技大学集成电路学部，陕西西安 710071
- 作者简介：
- 基金信息：
  
  Guangdong Basic and Applied Basic Research Foundation(2025A1515012058);Natural Science Basic Research Program of Shaanxi(2025JCYBQN822);Innovation Center of Radiation Application(KFZC2025010203);Shenzhen Science and Technology Program(KJZD20240903100506009);National Science Foundation of China(62090043;61934006);National Key Research and Development Program of China(2022YFB4401301)
- DOI：10.12263/DZXB.20260284
  CLC： TP183;TN791
- Received：21 April 2026，
  
  Accepted：01 June 2026，
  
  Published：25 June 2026
- 稿件说明：
移动端阅览
徐长卿, 莫桂棋, 刘毅, 等. FF-STDP：一种基于前向-前向与STDP融合的低功耗在线学习SNN加速器[J]. 电子学报, 2026, 54(06): 2443-2453.

XU Changqing, MO Guiqi, LIU Yi, et al. FF-STDP: A Forward-Forward STDP-Based Energy-Efficient SNN Accelerator with On-Chip Learning[J]. Acta Electronica Sinica, 2026, 54(06): 2443-2453.
徐长卿, 莫桂棋, 刘毅, 等. FF-STDP：一种基于前向-前向与STDP融合的低功耗在线学习SNN加速器[J]. 电子学报, 2026, 54(06): 2443-2453. DOI：10.12263/DZXB.20260284

XU Changqing, MO Guiqi, LIU Yi, et al. FF-STDP: A Forward-Forward STDP-Based Energy-Efficient SNN Accelerator with On-Chip Learning[J]. Acta Electronica Sinica, 2026, 54(06): 2443-2453. DOI：10.12263/DZXB.20260284

摘要

脉冲神经网络（Spiking Neural Network，SNN）作为第三代人工神经网络（Artificial Neural Network，ANN），通过离散脉冲事件进行信息编码与传递，具有事件驱动、时空信息融合以及低功耗计算等特点，在边缘智能、类脑计算和神经形态芯片等领域展现出广阔的应用前景。随着智能终端和物联网设备的快速发展，设备对环境自适应和持续学习能力的需求不断提升，在线学习逐渐成为边缘智能系统的重要研究方向。然而，现有在线学习方法主要分为两类：一类基于反向传播（BackPropagation，BP）机制，能够利用全局监督信息获得较高精度，但需存储大量中间状态和误差信息，导致计算复杂度、存储开销和能耗显著增加；另一类基于脉冲时序依赖可塑性（Spike-Timing-Dependent Plasticity，STDP）机制，具有局部更新和事件驱动等特点，易于硬件实现，但由于缺乏有效的监督信息，其学习能力和任务性能通常受到限制。针对上述问题，本文提出一种融合前向-前向（Forward-Forward，FF）学习范式与STDP机制的在线学习算法FF-STDP。该算法借鉴FF逐层局部优化思想，将教师信号引入脉冲学习过程，通过构建基于神经元脉冲发放率的局部优度（Goodness）评价机制，实现无需误差BP的半监督学习。在保留STDP局部权值更新特性的同时，引入FF提供的监督信息，使学习过程兼具较强的任务驱动能力与硬件友好性。针对传统FF算法在SNN中计算复杂、硬件实现成本较高的问题，本文采用脉冲计数近似梯度计算，并结合局部导数近似策略简化权值更新过程，使突触更新仅依赖于前后神经元的脉冲统计信息，显著降低计算复杂度和存储开销。基于所提出的FF-STDP算法，本文设计了一种面向边缘智能应用的事件驱动SNN加速器架构。该架构采用地址事件表示（Address Event Representation，AER）通信机制，结合并行突触后神经元处理、双口存储器和流水线执行策略，实现推理与学习的高效协同执行。为进一步降低硬件复杂度，设计了基于查找表（LookUp Table，LUT）的权值更新单元，实现低开销、高吞吐率的在线学习过程。为验证算法与硬件架构的有效性，本文在Xilinx ZYNQ-7010平台上完成系统实现，并采用混合国家标准与技术研究院数据库（Modified National Institute of Standards and Technology database，MNIST）和时尚-MNIST（Fashion-MNIST）数据集进行实验评估。实验结果表明：所提出的FF-STDP算法在MNIST和Fashion-MNIST数据集上的分类准确率分别达到90.43%和80.53%。在100 MHz工作频率下，加速器核心动态功耗仅为15 mW，推理速度达每秒1 887帧。与现有BP或传统STDP在线学习SNN硬件相比，所提方法在保持竞争性分类性能的同时，降低了硬件资源消耗和功耗开销，展现出良好的能效优势和边缘部署潜力。

Abstract

Spiking neural networks (SNNs)

regarded as the third generation of artificial neural networks (ANNs)

encode and transmit information through discrete spike events. Owing to their event-driven nature

spatiotemporal information processing capability

and low-power characteristics

SNNs have shown great potential in edge intelligence

brain-inspired computing

and neuromorphic hardware. With the rapid development of intelligent terminals and internet-of-things devices

the demand for environmental adaptation and continual learning is increasing

making online learning an important research direction for edge intelligence systems. Existing online learning methods can generally be categorized into two groups. Backpropagation (BP)-based methods achieve high accuracy by utilizing global supervisory information

but require the storage of a large number of intermediate states and error signals

resulting in considerable computational complexity

memory overhead

and power consumption. In contrast

spike-timing-dependent plasticity (STDP)-based methods adopt local and event-driven synaptic updates

making them hardware-friendly; however

their learning capability and task performance are often limited due to the lack of effective supervisory information. To address these challenges

this paper proposes an online learning algorithm

termed forward-forward (FF)-STDP

which combines the FF learning paradigm with the STDP mechanism. Inspired by the layer-wise local optimization strategy of FF

the proposed algorithm introduces teacher signals into the spike-based learning process and constructs a local goodness metric based on neuronal firing rates

thereby enabling supervised learning without error backpropagation. By preserving the local weight-update property of STDP while incorporating supervisory information provided by FF

the proposed method achieves both task-oriented learning capability and hardware efficiency. Furthermore

to reduce the computational complexity and hardware implementation cost of conventional FF algorithms in SNNs

spike-count-based gradient approximation and local derivative approximation are adopted

such that synaptic updates depend only on the spike statistics of pre-and post-synaptic neurons

significantly reducing computational and memory overhead.Based on the proposed FF-STDP algorithm

an event-driven SNN accelerator architecture for edge intelligence applications is developed. The architecture adopts an address event representation (AER) communication scheme and integrates parallel post-synaptic neuron processing

dual-port memory

and pipelined execution strategies to enable efficient collaborative execution of inference and learning. To further reduce hardware complexity

a lookup-table (LUT)-based weight update unit is designed

achieving low-overhead and high-throughput online learning. To validate the effectiveness of the proposed algorithm and hardware architecture

the system is implemented on a Xilinx ZYNQ-7010 platform and evaluated using the modified national institute of standards and technology database (MNIST) and Fashion-MNIST datasets. Experimental results demonstrate that the proposed FF-STDP algorithm achieves classification accuracies of 90.43% and 80.53% on MNIST and Fashion-MNIST

respectively. Operating at 100 MHz

the accelerator core consumes only 15 mW of dynamic power and achieves an inference throughput of 1 887 frames per second. Compared with existing BP-based and conventional STDP-based online learning SNN hardware

the proposed approach significantly reduces hardware resource consumption and power overhead while maintaining competitive classification performance

demonstrating excellent energy efficiency and potential for edge deployment.

关键词

Keywords

references

Maas W . Networks of spiking neurons: The third generation of neural network models [J ] . Transactions of the Society for Computer Simulation International , 1997 , 14 ( 4 ): 1659 - 1671 .

Han Song , Pool J , Tran J , et al . Learning both weights and connections for efficient neural networks [C ] // Proceedings of the 29th International Conference on Neural Information Processing Systems . Cambridge : MIT Press , 2015 : 1135 - 1143 .

Lobo J L , Del Ser J , Bifet A , et al . Spiking neural networks and online learning: An overview and perspectives [J ] . Neural Networks , 2020 , 121 : 88 - 100 . DOI: 10.1016/j.neunet.2019.09.004 http://dx.doi.org/10.1016/j.neunet.2019.09.004

Frenkel C , Lefebvre M , Legat J D , et al . A 0.086-mm 2 12.7-pJ/SOP 64k-synapse 256-neuron online-learning digital spiking neuromorphic processor in 28-nm CMOS [J ] . IEEE Transactions on Biomedical Circuits and Systems , 2019 , 13 ( 1 ): 145 - 158 .

Wang Qian , Li Youjie , Shao Botang , et al . Energy efficient parallel neuromorphic architectures with approximate arithmetic on FPGA [J ] . Neurocomputing , 2017 , 221 : 146 - 158 . DOI: 10.1016/j.neucom.2016.09.071 http://dx.doi.org/10.1016/j.neucom.2016.09.071

Wu Jiajun , Zhan Yi , Peng Zixuan , et al . Efficient design of spiking neural network with STDP learning based on fast CORDIC [J ] . IEEE Transactions on Circuits and Systems I: Regular Papers , 2021 , 68 ( 6 ): 2522 - 2534 . DOI: 10.1109/tcsi.2021.3061766 http://dx.doi.org/10.1109/tcsi.2021.3061766

Li Sixu , Zhang Zhaomin , Mao Ruixin , et al . A fast and energy-efficient SNN processor with adaptive clock/event-driven computation scheme and online learning [J ] . IEEE Transactions on Circuits and Systems I: Regular Papers , 2021 , 68 ( 4 ): 1543 - 1552 . DOI: 10.1109/tcsi.2021.3052885 http://dx.doi.org/10.1109/tcsi.2021.3052885

He Zhen , Shi Cong , Wang Tengxiao , et al . A low-cost FPGA implementation of spiking extreme learning machine with on-chip reward-modulated STDP learning [J ] . IEEE Transactions on Circuits and Systems II: Express Briefs , 2022 , 69 ( 3 ): 1657 - 1661 . DOI: 10.1109/tcsii.2021.3117699 http://dx.doi.org/10.1109/tcsii.2021.3117699

Zhang Jian , Wang Ran , Pei Xudong , et al . A fast spiking neural network accelerator based on BP-STDP algorithm and weighted neuron model [J ] . IEEE Transactions on Circuits and Systems II: Express Briefs , 2022 , 69 ( 4 ): 2271 - 2275 . DOI: 10.1109/tcsii.2021.3137987 http://dx.doi.org/10.1109/tcsii.2021.3137987

Liu Wei , Xiao Shanlin , Liu Yue , et al . SC-PLR: An approximate spiking neural network accelerator with on-chip predictive learning rule [J ] . IEEE Transactions on Biomedical Circuits and Systems , 2024 , 18 ( 5 ): 1156 - 1165 . DOI: 10.1109/tbcas.2024.3385235 http://dx.doi.org/10.1109/tbcas.2024.3385235

Shi Cong , Fu Xiang , Wang Haibing , et al . Ghost reservoir: A memory-efficient low-power and real-time neuromorphic processor of liquid state machine with on-chip learning [J ] . IEEE Transactions on Circuits and Systems II: Express Briefs , 2024 , 71 ( 10 ): 4526 - 4530 . DOI: 10.1109/tcsii.2024.3395415 http://dx.doi.org/10.1109/tcsii.2024.3395415

Xu Changqing , Liu Yi , Chen Dongdong , et al . Direct training via backpropagation for ultra-low-latency spiking neural networks with multi-threshold [J ] . Symmetry , 2022 , 14 ( 9 ): 1933 . DOI: 10.3390/sym14091933 http://dx.doi.org/10.3390/sym14091933

Xu Changqing , Zhang Wenrui , Liu Yu , et al . Boosting throughput and efficiency of hardware spiking neural accelerators using time compression supporting multiple spike codes [J ] . Frontiers in Neuroscience , 2020 , 14 : 104 . DOI: 10.3389/fnins.2020.00104 http://dx.doi.org/10.3389/fnins.2020.00104

Xu Changqing , Liu Yi , Yang Yintang . STCSNN: High energy efficiency spike-train level spiking neural networks with spatio-temporal conversion [J ] . Neurocomputing , 2024 , 607 : 128364 . DOI: 10.1016/j.neucom.2024.128364 http://dx.doi.org/10.1016/j.neucom.2024.128364

Lee J , Zhang Renqian , Zhang Wenrui , et al . Spike-train level direct feedback alignment: Sidestepping backpropagation for on-chip training of spiking neural nets [J ] . Frontiers in Neuroscience , 2020 , 14 : 143 . DOI: 10.3389/fnins.2020.00143 http://dx.doi.org/10.3389/fnins.2020.00143

Hinton G . The forward-forward algorithm: Some preliminary investigations [PP/OL ] . V1. arXiv ( 2022-12-27 )[ 2026-03-01 ] . https://arxiv.org/abs/2212.13345 https://arxiv.org/abs/2212.13345 .

Fang Wei , Chen Yanqi , Ding Jianhao , et al . SpikingJelly: An open-source machine learning infrastructure platform for spike-based intelligence [J ] . Science Advances , 2023 , 9 ( 40 ): eadi1480 . DOI: 10.1126/sciadv.adi1480 http://dx.doi.org/10.1126/sciadv.adi1480

Tavanaei A , Maida A . BP-STDP: Approximating backpropagation using spike timing dependent plasticity [J ] . Neurocomputing , 2019 , 330 : 39 - 47 . DOI: 10.1016/j.neucom.2018.11.014 http://dx.doi.org/10.1016/j.neucom.2018.11.014

Bu Tong , Fang Wei , Ding Jianhao , et al . Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks [PP/OL ] . V1. arXiv ( 2023-03-08 )[ 2026-03-01 ] . https://arxiv.org/abs/2303.04347 https://arxiv.org/abs/2303.04347 .

Lecun Y , Bottou L , Bengio Y , et al . Gradient-based learning applied to document recognition [J ] . Proceedings of the IEEE , 1998 , 86 ( 11 ): 2278 - 2324 . DOI: 10.1109/5.726791 http://dx.doi.org/10.1109/5.726791

Xiao Han , Rasul K , Vollgraf R . Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms [PP/OL ] . V2. arXiv ( 2017-09-15 )[ 2026-03-01 ] . https://arxiv.org/abs/1708.07747 https://arxiv.org/abs/1708.07747 .

Orchard G , Jayawant A , Cohen G K , et al . Converting static image datasets to spiking neuromorphic datasets using saccades [J ] . Frontiers in Neuroscience , 2015 , 9 : 437 . DOI: 10.3389/fnins.2015.00437 http://dx.doi.org/10.3389/fnins.2015.00437

Amir A , Taba B , Berg D , et al . A low power, fully event-based gesture recognition system [C ] // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . Piscataway : IEEE , 2017 : 7388 - 7397 . DOI: 10.1109/cvpr.2017.781 http://dx.doi.org/10.1109/cvpr.2017.781

Krizhevsky A . Learning multiple layers of features from tiny images [R ] . Toronto : University of Toronto , 2009 .

Cheng Xi , Cao Shu , Wang Shangmei , et al . An FPGA-based event-driven SNN accelerator for DVS applications with structured sparsity and early-stop [J ] . IEEE Transactions on Circuits and Systems I: Regular Papers , 2025 , 72 ( 7 ): 3298 - 3310 . DOI: 10.1109/tcsi.2025.3560666 http://dx.doi.org/10.1109/tcsi.2025.3560666

Ma Jingxiao , Panda P , Reda S . FF-INT8: Efficient forward-forward DNN training on edge devices with INT8 precision [C ] // Proceedings of the 62nd ACM/IEEE Design Automation Conference (DAC) . Piscataway : IEEE , 2025 : 1 - 7 . DOI: 10.1109/dac63849.2025.11133061 http://dx.doi.org/10.1109/dac63849.2025.11133061

王莹 , 高岚 , 张哲 , 等 . 基于非一般类算子融合方法及硬件架构设计 [J ] . 电子学报 , 2025 , 53 ( 9 ): 3299 - 3309 .

Wang Ying , Gao Lan , Zhang Zhe , et al . Operator fusion method and hardware architecture design based on non-standard operators [J ] . Acta Electronica Sinica , 2025 , 53 ( 9 ): 3299 - 3309 . (in Chinese)

Views

149

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

A Convolutional Neural Network Accelerator Based on Inter-Layer Feature Map Compression

An Emergency Multipath Transmission Method Based on Online Learning Adaptive Optimal Selection in Disaster Scenarios

Gradient Re-Weighting Guided by Feature Fusion for Online Continual Learning

Concept Evolution Detection Method for Open Feature Space

Multi-Base-Station Cooperative Integrated Sensing and Communication Design for Nearshore Scenarios

Related Author

BU Wenjie

ZHOU Mu

CHEN Lin

XIE Liangbo

YANG Jinrui

LI Kangrui

XU Changqiao

WANG Mu

Related Institution

School of Electronic Science and Engineering, Chongqing University of Posts and Telecommunications

School of Communications and Information Engineering, Chongqing University of Posts and Telecommunications

State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications

School of Information and Communication Engineering, University of Electronic Science and Technology of China

School of Computer and Information Technology, Shanxi University

⁰