Deep Neural Networks have demonstrated impressive capabilities across various domains, yet their inherent complexity often obscures the rationale behind their predictions. This opacity poses challenges in domains where explainability is critical. Here, we present a novel methodology inspired by signal processing that leverages Singular Value Decomposition to both remove the redundancy in the neural network and derive compressed feature representations to be analyzed with clustering. We carried out empirical experiments with a network of the VGG family trained on CIFAR-10 and FMNIST datasets, and propose two strategies to address the trustworthiness issue in AI decisions.
Manovi, L., Capelli, L., Marchioni, A., Martinini, F., Setti, G., Mangia, M., et al. (2024). SVD-based Peephole and Clustering to Enhance Trustworthiness in DNN Classifiers. 345 E 47TH ST, NEW YORK, NY 10017 USA : IEEE [10.1109/aicas59952.2024.10595919].
SVD-based Peephole and Clustering to Enhance Trustworthiness in DNN Classifiers
Manovi, LiviaPrimo
;Marchioni, Alex;Martinini, Filippo;Setti, Gianluca;Mangia, Mauro;Rovatti, Riccardo
2024
Abstract
Deep Neural Networks have demonstrated impressive capabilities across various domains, yet their inherent complexity often obscures the rationale behind their predictions. This opacity poses challenges in domains where explainability is critical. Here, we present a novel methodology inspired by signal processing that leverages Singular Value Decomposition to both remove the redundancy in the neural network and derive compressed feature representations to be analyzed with clustering. We carried out empirical experiments with a network of the VGG family trained on CIFAR-10 and FMNIST datasets, and propose two strategies to address the trustworthiness issue in AI decisions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
