Deep Learning Reliability: Towards Mitigating Reliability Threats in Deep Learning Systems by Exploiting Intrinsic Characteristics of DNNs

Muhammad Abdullah Hanif, Muhammad Shafique

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract

Deep Learning (DL) models have shown remarkable performance for a variety of AI applications such as image classification, object detection, semantic segmentation, and natural language processing. Due to their state-of-the-art performance, DL models are also being used in safety-critical applications, e.g., autonomous driving and smart healthcare. However, these applications have strict robustness constraints which have to be satisfied to meet the industrial standards. Therefore, the robustness of these DL models against different reliability threats has to be thoroughly studied and methods have to be designed to improve their inherent resilience against all types of reliability threats. Towards this, this chapter first highlights different types of reliability threats that can significantly degrade the performance of a DL system. The chapter also highlight different error-resilience characteristics of DL models that can be exploited to design low-cost fault-mitigation techniques. The chapter then covers different techniques for mitigating hardware-induced reliability threats in DL systems without incurring high design-time and run-time overheads.

Original languageEnglish (US)
Title of host publicationEmbedded Machine Learning for Cyber-Physical, IoT, and Edge Computing
Subtitle of host publicationUse Cases and Emerging Challenges
PublisherSpringer Nature
Pages553-568
Number of pages16
ISBN (Electronic)9783031406775
ISBN (Print)9783031406768
DOIs
StatePublished - Jan 1 2023

Keywords

  • Aging
  • Deep learning
  • DNN reliability
  • Error resilience
  • Fault mitigation
  • Neural networks
  • Permanent faults
  • Robustness
  • Soft errors
  • Timing errors

ASJC Scopus subject areas

  • General Computer Science
  • General Engineering
  • General Social Sciences

Fingerprint

Dive into the research topics of 'Deep Learning Reliability: Towards Mitigating Reliability Threats in Deep Learning Systems by Exploiting Intrinsic Characteristics of DNNs'. Together they form a unique fingerprint.

Cite this