Anouar Chahbouni, Khaoula El Manaa, Imane El Manaa, Yassine Abouch, Badre Bossoufi, Mohammed El Ghzaoui, Rachid El Alami · Computers 2026 · 2026
DOI: 10.3390/computers15110812
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Deep neural networks achieve remarkable performance in visual recognition tasks, but this accuracy often comes at the cost of high computational complexity and over-parameterized representations. Structured pruning has emerged as an effective solution for reducing this complexity by eliminating unnecessary representations. Nevertheless, its success implicitly relies on the existence of exploitable redundancy within learned feature spaces, whereas the nature, structure, and evolution of such redundancy remain largely underexplored. To address this gap, this paper presents a comprehensive redundancy analysis framework that investigates deep representations from complementary visual, geometric, and intrinsic dimensionality perspectives. The framework is evaluated using VGG16 and Swin-Tiny on the CIFAR-10 and PlantVillage datasets, enabling the analysis of redundancy across convolutional and transformer-based architectures, multiple representation levels, and distinct image classification domains. The proposed framework combines feature-map visualization, cosine similarity analysis, singular value decomposition (SVD), and effective rank (ER) to jointly characterize feature redundancy, representational dependencies, and intrinsic dimensionality from complementary local and global perspectives. The results reveal that redundancy is a pervasive yet dynamic property of deep neural networks, evolving according to architecture, dataset characteristics, representation depth, and input distributions. Pairwise redundancy ratios reach up to 18.61% in deep representations, whereas normalized effective rank values decrease to as low as 0.01, indicating strong spectral concentration in some representations. More importantly, the study demonstrates that redundancy extends beyond directly similar features and also encompasses global dependencies and representations occupying reduced effective subspaces. Structured pruning validation further demonstrates substantial compression with preserved predictive performance under the evaluated experimental settings. These findings provide a deeper understanding of learned representations, establish a principled perspective for interpreting pruning effectiveness, and offer practical guidelines for developing future redundancy-aware compression methods for efficient and deployable deep learning systems.
No comments yet — start the discussion below.