ORCID
- Vasilios Kelefouras: 0000-0001-9591-913X
Abstract
Convolutional Neural Networks (CNNs) achievestate-of-the-art performance across various application domainsbut are often resource-intensive, limiting their use on resourceconstrained devices. Low-rank factorization (LRF) has emerged asa promising technique to reduce the computational complexity andmemory footprint of CNNs, enabling efficient deployment withoutsignificant performance loss. However, challenges still remainin optimizing the rank selection problem, balancing memoryreduction and accuracy, and integrating LRF into the trainingprocess of CNNs. In this paper, a novel and generic methodologyfor layer-wise rank selection is presented, considering inter-layerinteractions. Our approach is compatible with any decompositionmethod and does not require additional retraining. The proposedmethodology is evaluated in thirteen widely-used, CNN models,significantly reducing model parameters and Floating-Point Operations (FLOPs). In particular, our approach achieves up to a94.6% parameter reduction (82.3% on average) and up to 90.7%FLOPs reduction (59.6% on average), with less than a 1.5% dropin validation accuracy, demonstrating superior performance andscalability compared to existing techniques.
Publication Date
2025-04-02
Event
DATE 25 (Design, Automation and Test in Europe)
Publication Title
DATE 2025 Conference
Acceptance Date
2025-01-20
Deposit Date
2025-01-17
Embargo Period
2025-04-02
Additional Links
Recommended Citation
Kokhazadeh, M., Keramidas, G., Kelefouras, V., & Stamoulis, I. (2025) 'A CNN Compression Methodology for Layer-Wise Rank Selection Considering Inter-Layer Interactions', DATE 2025 Conference, . Retrieved from https://pearl.plymouth.ac.uk/secam-research/2071
