Are CNNs reliable enough for critical applications? An exploratory study

Mohamed Ayoub Neggaz (1) , Ihsen Alouani (2, 3, 4, 5) , Smail Niar (1) , Fadi Kurdahi (6)

1 LAMIH - Laboratoire d'Automatique, de Mécanique et d'Informatique industrielles et Humaines - UMR 8201
2 IEMN - Institut d’Électronique, de Microélectronique et de Nanotechnologie - UMR 8520
3 COMNUM - IEMN - COMmunications NUMériques - IEMN
4 IEMN-DOAE - Institut d’Électronique, de Microélectronique et de Nanotechnologie - Département Opto-Acousto-Électronique - UMR 8520
5 INSA Hauts-De-France - INSA Institut National des Sciences Appliquées Hauts-de-France
6 CECs - Centro de Estudios Científicos

Resource-constrained CNN implementations are subject to various reliability threats. This article provides an exploratory study investigating the impact of faults (soft errors modeled as bit flips) across the parameters of CNNs and the impact on the CNN weights, which can be used to create reliability-aware design guidelines. -Theocharis Theocharides, University of Cyprus -Muhammad Shafique, Technische Universitat Wien

Mots clés: Engines; Fault tolerant systems; Reliability engineering; Quantization (signal); Feature extraction; Machine Learning; Convolutional Neural Networks; Reliability; Single Event Upsets; Multiple Event Upsets; Embedded Systems

Domaines: Sciences de l'ingénieur [physics]; Informatique [cs]; Intelligence artificielle [cs.AI]; Réseaux et télécommunications [cs.NI]; Traitement du signal et de l'image [eess.SP]; Electronique

Mohamed Ayoub Neggaz, Ihsen Alouani, Smail Niar, Fadi Kurdahi. Are CNNs reliable enough for critical applications? An exploratory study. IEEE Design & Test, 2020, 37 (2), pp.76-83. ⟨10.1109/MDAT.2019.2952336⟩. ⟨hal-03321520⟩