Performance Optimization and Interpretability of Recurrent Sigma-Pi-Sigma Neural Networks on Application of IoE Data

With the exponential growth of Internet of Things (IoT) devices in the era of Internet of Everything (IoE), two major issues arise: 1) data processing speed and 2) interpretability in neural networks. Specifically, training neural networks to handle IoE data often results in poor convergence speed,...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:IEEE internet of things journal Ročník 12; číslo 4; s. 3639 - 3653
Hlavní autori: Deng, Fei, Zhang, Liqing
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Piscataway IEEE 15.02.2025
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Predmet:
Accuracy
Adaptation models
Adaptive algorithms
Adaptive momentum
Automation
Classification algorithms
Convergence
Data processing
Error functions
Heuristic algorithms
Internet of Everything (IoE)
Internet of Things
interpretability
Neural networks
Optimization
recurrent sigma-pi-sigma neural network (RSPSNN)
Regularization
smoothing lasso regularization
Smoothing methods
Training
Vectors
ISSN:2327-4662, 2327-4662
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:With the exponential growth of Internet of Things (IoT) devices in the era of Internet of Everything (IoE), two major issues arise: 1) data processing speed and 2) interpretability in neural networks. Specifically, training neural networks to handle IoE data often results in poor convergence speed, overfitting of the weights, and fluctuations in the error function. To address these challenges, this article introduces a novel neural network, the recurrent sigma-pi-sigma neural network (RSPSNN), trained using a batch gradient algorithm enhanced with smoothing L1 lasso regularization and an adaptive momentum term. This approach not only improves convergence speed but also enhances generalization capabilities and reduces oscillations. Furthermore, the interpretability of RSPSNN is theoretically demonstrated through characteristics of monotonicity, strong/weak convergence, and stability. Finally, the theoretical findings are supported by experiment results in classification, recognition, and prediction tasks.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2327-4662
2327-4662
DOI:10.1109/JIOT.2024.3472052