.png){kind=icon}
Return
Lightweight convolutional neural network based on overlapping kernel compression pruning for modeling nonlinear systems
.png){kind=icon}
PRE
AI
DOI：10.1007/s11431-026-3359-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Convolutional neural networks (CNNs) stack numerous convolutional kernels to model nonlinear systems by filtering spatial features. However, overlapping kernels with similar spatial weight arrangements introduce spurious correlations from repeated feature extraction, which results in high computational cost and limited generalization capability. To address this issue, a lightweight CNN (LCNN) based on overlapping kernel compression pruning is proposed in this paper. First, an overlapping kernel perception (OKP) method is designed to assess the kernel redundancy of LCNN. The OKP method adopts spectral clustering to group similar kernels into the graph Laplacian spectral space. It broadens the redundancy scope by incorporating overlapping kernels with similar spatial weight arrangements. Second, a kernel compression pruning (KCP) strategy is built to prune redundant kernels. Guided by kernel redundancy, the KCP strategy performs singular value decomposition on each group of similar kernels and adaptively truncates the rank of the decomposed matrix to reconstruct a compact set of representative kernels. Third, a regularized gradient descent algorithm is developed to fine-tune the parameters of LCNN after pruning. Finally, experimental results on three industrial datasets confirm the superior generalization performance of LCNN compared to other models.
Keywords:
convolutional neural networks
kernel compression pruning
generalization
kernel redundancy
Journal
IF:
4.9
Papers:
5.0K
Citations:
9.9K
Organization
No organization information available
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSave
View more