CoChat

Lead federated neuromorphic learning for wireless edge artificial intelligence

Open access
Yang, Helin, Lam, Kwok-Yan, Xiao, Liang, Xiong, Zehui, Hu, Hao, Niyato, Dusit, Vincent Poor, H. · 2022 · Nature Communications

Published version of a preprint

The DOI matches its Crossref record; 52 of the 52 references it cites were found in a registry or an index.

Abstract

In order to realize the full potential of wireless edge artificial intelligence (AI), very large and diverse datasets will often be required for energy-demanding model training on resource-constrained edge devices. This paper proposes a lead federated neuromorphic learning (LFNL) technique, which is a decentralized energy-efficient brain-inspired computing method based on spiking neural networks. The proposed technique will enable edge devices to exploit brain-like biophysiological structure to collaboratively train a global model while helping preserve privacy. Experimental results show that, under the situation of uneven dataset distribution among edge devices, LFNL achieves a comparable recognition accuracy to existing edge AI techniques, while substantially reducing data traffic by >3.5× and computational latency by >2.0×. Furthermore, LFNL significantly reduces energy consumption by >4.5× compared to standard federated learning with a slight accuracy loss up to 1.5%. Therefore, the proposed LFNL can facilitate the development of brain-inspired computing and edge AI.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-022-32020-w

Cites

52 of 52 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Schematic diagram of the lead federated neuromorphic learning system. a Schematic of a social learning network, where each human uses five sensory organs to interact with the outside environment via neural networks. The humans in a group exchange learned knowledge with each other for better recognition. b Inspired by the collaborative learning system in (a) a federated neuromorphic learning system is introduced to perform model aggregation from edge devices in a group. The integration between devices and the external environment is handled by using sensors (e.g., cameras, microphones, radars, and touch sensors). c The structure of an SNN which is adopted to perform neuromorphic computing for edge devices. d The principle of LFNL without a central server, where one device is selected as a leader to manage model aggregation in a group. e An example situation of multiple humans crossing a vehicular road, where multiple edge devices can observe, hear, and sense traffic objects. f Illustration of LFNL-based traffic recognition. The leader (device) leads other followers (devices) to train their own local neuromorphic models independently, and it collects local model parameters (w2, w3, …, wK) to perform model aggregation before broadcasting w to the followers for the next local training. The exchange of local and global parameters repeats until convergence.
Fig. 1 Schematic diagram of the lead federated neuromorphic learning system. a Schematic of a social learning network, where each human uses five sensory organs to interact with the outside environment via neural networks. The humans in a group exchange learned knowledge with each other for better recognition. b Inspired by the collaborative learning system in (a) a federated neuromorphic learning system is introduced to perform model aggregation from edge devices in a group. The integration between devices and the external environment is handled by using sensors (e.g., cameras, microphones, radars, and touch sensors). c The structure of an SNN which is adopted to perform neuromorphic computing for edge devices. d The principle of LFNL without a central server, where one device is selected as a leader to manage model aggregation in a group. e An example situation of multiple humans crossing a vehicular road, where multiple edge devices can observe, hear, and sense traffic objects. f Illustration of LFNL-based traffic recognition. The leader (device) leads other followers (devices) to train their own local neuromorphic models independently, and it collects local model parameters (w2, w3, …, wK) to perform model aggregation before broadcasting w to the followers for the next local training. The exchange of local and global parameters repeats until convergence.CC BY · PubMed Central
Fig. 2 LFNL for audio recognition on a road. a Principle of local neuromorphic learning (LNL) with data and model on the devices. b Principle of centralized neuromorphic learning (CNL) with data and model being stored at the central server. c Principle of centralized federated neuromorphic learning (CFNL) with data being kept on the devices, and local model parameters being uploaded to the central server for model aggregation. d Principle of transfer neuromorphic learning (TNL) with data being kept on the device side, and each device trains its model and then passes it to the next device for training, cyclically repeating the process. e An example of audio recognition on a road, including firetruck, ambulance and general traffic sounds. f, g Validation loss curves for three locally training devices and LFNL. h–j Box plots show test accuracy performed for three locally training devices and LFNL with uneven distributions of training dataset (F: firetruck sound class, A: ambulance sound class, T: general traffic sound class). The training dataset distributions of three sound classes for three devices are shown at the top of these figures. k Confusion matrix for the test set in LFNL after training. l–n Test accuracy, data traffic, and training latency comparisons for different learning methods, respectively.
Fig. 2 LFNL for audio recognition on a road. a Principle of local neuromorphic learning (LNL) with data and model on the devices. b Principle of centralized neuromorphic learning (CNL) with data and model being stored at the central server. c Principle of centralized federated neuromorphic learning (CFNL) with data being kept on the devices, and local model parameters being uploaded to the central server for model aggregation. d Principle of transfer neuromorphic learning (TNL) with data being kept on the device side, and each device trains its model and then passes it to the next device for training, cyclically repeating the process. e An example of audio recognition on a road, including firetruck, ambulance and general traffic sounds. f, g Validation loss curves for three locally training devices and LFNL. h–j Box plots show test accuracy performed for three locally training devices and LFNL with uneven distributions of training dataset (F: firetruck sound class, A: ambulance sound class, T: general traffic sound class). The training dataset distributions of three sound classes for three devices are shown at the top of these figures. k Confusion matrix for the test set in LFNL after training. l–n Test accuracy, data traffic, and training latency comparisons for different learning methods, respectively.CC BY · PubMed Central
Fig. 3 LFNL for visual recognition on a road. a An example of visual recognition (image classification) on a road, including bicycle, car, and traffic light image classes. b Validation loss curves for three locally training devices and LFNL. c Confusion matrix for the test set in LFNL after training. d–f Box plots show test accuracy performed for three locally training devices and LFNL with uneven distributions of training dataset (B: bicycle image class, C: car image class, L: traffic light image class). g Illustration of box plots showing test accuracy performed for six locally training devices and LFNL. h–j Validation loss curves, test accuracy and training latency performance comparisons for different learning methods, respectively.
Fig. 3 LFNL for visual recognition on a road. a An example of visual recognition (image classification) on a road, including bicycle, car, and traffic light image classes. b Validation loss curves for three locally training devices and LFNL. c Confusion matrix for the test set in LFNL after training. d–f Box plots show test accuracy performed for three locally training devices and LFNL with uneven distributions of training dataset (B: bicycle image class, C: car image class, L: traffic light image class). g Illustration of box plots showing test accuracy performed for six locally training devices and LFNL. h–j Validation loss curves, test accuracy and training latency performance comparisons for different learning methods, respectively.CC BY · PubMed Central
Fig. 4 LFNL for radar signal recognition on a road. a An example of radar signal recognition on a road, which has five radar gesture classes. b Validation loss curves for three locally training devices and LFNL. c, d Box plots show test accuracy and the histogram shows the average test accuracy performed for three locally training devices and LFNL after training. e Illustration of box plots showing test accuracy performed for six locally training devices and LFNL, where the training dataset is divided into six smaller parts for them. f Confusion matrix for the test set in LFNL after training. g, h Test accuracy and training latency performance comparisons for different methods, respectively.
Fig. 4 LFNL for radar signal recognition on a road. a An example of radar signal recognition on a road, which has five radar gesture classes. b Validation loss curves for three locally training devices and LFNL. c, d Box plots show test accuracy and the histogram shows the average test accuracy performed for three locally training devices and LFNL after training. e Illustration of box plots showing test accuracy performed for six locally training devices and LFNL, where the training dataset is divided into six smaller parts for them. f Confusion matrix for the test set in LFNL after training. g, h Test accuracy and training latency performance comparisons for different methods, respectively.CC BY · PubMed Central
Fig. 5 Recognition accuracy and computation energy comparisons between ANNs and SNNs trained on the audio, vision and radar datasets. a Audio classification accuracy comparison between LFNL-SNN and LFL-ANN versus the spike train duration T of the operation of SNNs on the traffic sound dataset. Both SNNs and ANNs have the same learning structure with 128-500-3 neurons (for input-hidden-label layers). b, c Average test accuracy and energy consumption comparisons when the spike train duration T is 15. d Visual classification accuracy comparison between LFNL-SNN and LFL-ANN versus T of SNNs on traffic image dataset. Both SNNs and ANNs have the same learning structure with 1728-2500-3 neurons. e, f Average test accuracy and energy consumption comparisons when the spike train duration T is 15. g Radar gesture classification accuracy comparison between LFNL-SNN and LFL-ANN versus T of SNNs on radar gesture dataset. Both SNNs and ANNs have the same structure with 4800-1000-5 neurons. h, i Average test accuracy and energy consumption comparisons when the spike train duration T is 15.
Fig. 5 Recognition accuracy and computation energy comparisons between ANNs and SNNs trained on the audio, vision and radar datasets. a Audio classification accuracy comparison between LFNL-SNN and LFL-ANN versus the spike train duration T of the operation of SNNs on the traffic sound dataset. Both SNNs and ANNs have the same learning structure with 128-500-3 neurons (for input-hidden-label layers). b, c Average test accuracy and energy consumption comparisons when the spike train duration T is 15. d Visual classification accuracy comparison between LFNL-SNN and LFL-ANN versus T of SNNs on traffic image dataset. Both SNNs and ANNs have the same learning structure with 1728-2500-3 neurons. e, f Average test accuracy and energy consumption comparisons when the spike train duration T is 15. g Radar gesture classification accuracy comparison between LFNL-SNN and LFL-ANN versus T of SNNs on radar gesture dataset. Both SNNs and ANNs have the same structure with 4800-1000-5 neurons. h, i Average test accuracy and energy consumption comparisons when the spike train duration T is 15.CC BY · PubMed Central
Fig. 6 Classification accuracy evaluation on CIFAR10 and CIFAR100 datasets. a, b Performance evaluation of LFNL vs CFNL under different device configurations. c, d Impact of the non-IID data factor on the classification accuracy of LFNL vs CFNL when the training dataset is divided among 100 devices and 30 devices participating in each global round. e, f Effects of the probability of stragglers on the performance of LFNL vs CFNL when the training dataset is divided among 100 devices and 30 devices participating in each global round.
Fig. 6 Classification accuracy evaluation on CIFAR10 and CIFAR100 datasets. a, b Performance evaluation of LFNL vs CFNL under different device configurations. c, d Impact of the non-IID data factor on the classification accuracy of LFNL vs CFNL when the training dataset is divided among 100 devices and 30 devices participating in each global round. e, f Effects of the probability of stragglers on the performance of LFNL vs CFNL when the training dataset is divided among 100 devices and 30 devices participating in each global round.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 52 references

Research works printed but not linked: 2

  1. Che, C., Li, X., Chen, C., He, X. & Zheng, Z. A decentralized federated learning framework via committee mechanism with convergence guarantee. Preprint at https://arxiv.org/pdf/2108.00365.pdf (2021). · arXiv
  2. Chintamani, N., Yash, M., Nikhil, P., Vivek, P. & Sandeep, P. Smart traffic control using deep learning. Preprint at https://github.com/nikhilpatil99/Smart-Traffic-Management-Using-Deep-Learning (2019).

Cited by

See all 143

Metadata

Authors
Yang, H., Lam, K.-Y., Xiao, L., Xiong, Z., Hu, H., Niyato, D., & Vincent Poor, H.
Affiliations
  • Nanyang Technological University (SG)
  • Xiamen University (CN)
  • Singapore University of Technology and Design (SG)
Corresponding author
Lam, Kwok-Yan
Venue
Nature Communications
Published
2022-07-25
Volume
13
Issue
1
Pages
4269-4269
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4287218656
Cited by
127
References
47
Citations by year
2026: 22 · 2025: 39 · 2024: 28 · 2023: 30 · 2022: 8
FWCI
8.64
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DR-NTU (Nanyang Technological University) (submittedVersion, cc-by)
Funders
  • National Science Foundation (1908308)
  • National Research Foundation (SRG-ISTD-2021-165)
  • National Research Foundation Singapore
  • Nanyang Technological University (RG16/20, Tier 1)
  • National Natural Science Foundation of China (1908308, 61971366, U21A20444, SUTD SRG-ISTD-2021-165)
  • Ministry of Education, India
  • Singapore University of Technology and Design (SRG-ISTD-2021165)
  • Info-communications Media Development Authority
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
Keywords
federated learning, spiking neural networks, neuromorphic computing, energy efficiency, privacy preservation, non-IID data, data traffic reduction, edge devices
MeSH terms
Artificial Intelligence, Brain, Neural Networks, Computer, Recognition, Psychology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Lead federated neuromorphic learning for wireless edge artificial intelligence. https://app.cochat.ai/papers/10.1038/s41467-022-32020-w
The paper
Yang, H., Lam, K.-Y., Xiao, L., Xiong, Z., Hu, H., Niyato, D., & Vincent Poor, H. (2022). Lead federated neuromorphic learning for wireless edge artificial intelligence. Nature Communications. https://doi.org/10.1038/s41467-022-32020-w

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.