CoChat

Artificial Neural Networks for Space and Safety-Critical Applications: Reliability Issues and Potential Solutions

Open access
Rech, Paolo · 2024 · IEEE Transactions on Nuclear Science

The DOI matches its Crossref record; 165 of the 165 references it cites were found in a registry or an index.

Abstract

Machine Learning is among the greatest advancements in computer science and engineering and is today used to classify or detect objects, a key feature in autonomous vehicles. Since neural networks are heavily used in safety-critical applications, such as automotive and aerospace, their reliability must be paramount. However, the reliability evaluation of neural network systems is extremely challenging due to the complexity of the software, which is composed of hundreds of layers, and of the underlying hardware, typically a parallel device or an embedded accelerator. This paper reviews fundamental concepts of Artificial Intelligence, Deep Neural Network, and parallel computing device reliability. Then, the reliability studies that consider the radiation effects in the hardware, their propagation through the computing architecture, and their final impact on the software output are summarized. A detailed survey of the available strategies to measure the sensitivity of neural network frameworks and to observe fault propagation is given, together with a summary of the data obtained so far. Finally, a discussion on how to use the experimental evaluation to design effective and efficient hardening solutions for Artificial Neural Networks is provided. The available hardening solutions are critically reviewed, highlighting their benefits and drawbacks.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1109/tns.2024.3349956

Cites

165 of 165 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Classification of Artificial Intelligence, Machine Learning, and Deep Learning.
Fig. 1 Classification of Artificial Intelligence, Machine Learning, and Deep Learning.CC-BY-NC-ND · From the PDF
Fig. 2 Visualization of the structure of a Convolutional Neural Network. The Convolution layers extract features from the input, ReLU operations ensure nonlinear input-output correlation, Pooling layers reduce the amount of data propagated, the fully connected layers use the output of previous layers to detect and classify objects. Taken from [20].
Fig. 2 Visualization of the structure of a Convolutional Neural Network. The Convolution layers extract features from the input, ReLU operations ensure nonlinear input-output correlation, Pooling layers reduce the amount of data propagated, the fully connected layers use the output of previous layers to detect and classify objects. Taken from [20].CC-BY-NC-ND · From the PDF
Fig. 3 Example of the convolution operation. The convolution needs to be performed in every window of the input feature map and, thus, is highly computationally demanding. Adapted from [20].
Fig. 3 Example of the convolution operation. The convolution needs to be performed in every window of the input feature map and, thus, is highly computationally demanding. Adapted from [20].CC-BY-NC-ND · From the PDF
Fig. 4 Example of features extracted by the convolution layers from the input image and of the effect of ReLu and Pooling. Taken from [24].
Fig. 4 Example of features extracted by the convolution layers from the input image and of the effect of ReLu and Pooling. Taken from [24].CC-BY-NC-ND · From the PDF
Fig. 5 Example of max and average pooling. Max pooling propagates only the element with the highest value while average pooling propagates the average value of the elements in the window.
Fig. 5 Example of max and average pooling. Max pooling propagates only the element with the highest value while average pooling propagates the average value of the elements in the window.CC-BY-NC-ND · From the PDF
Fig. 6 Simplified view of the architecture of modern Graphics Processing Units, that are composed of an array of Streaming Multiprocessors that share L2 cache. Each SM has several computing cores with different precision. The view is based on NVIDIA CUDA devices, but can be applied to any GPU architecture.
Fig. 6 Simplified view of the architecture of modern Graphics Processing Units, that are composed of an array of Streaming Multiprocessors that share L2 cache. Each SM has several computing cores with different precision. The view is based on NVIDIA CUDA devices, but can be applied to any GPU architecture.CC-BY-NC-ND · From the PDF
Fig. 7 Simplified view of the architecture of FPGAs, showing the programmable elements used to define a circuit to be executed [39].
Fig. 7 Simplified view of the architecture of FPGAs, showing the programmable elements used to define a circuit to be executed [39].CC-BY-NC-ND · From the PDF
Image 8
CC-BY-NC-ND · From the PDF
Fig. 9 High level schematic of the Coral Edge TPU architecture, which is basically composed of a hardwired systolic array. Adapted from [5].
Fig. 9 High level schematic of the Coral Edge TPU architecture, which is basically composed of a hardwired systolic array. Adapted from [5].CC-BY-NC-ND · From the PDF
Fig. 10 Possible effect of a single event effect corrupting critical or shared resources of a parallel device. As a reult of the impact, multiple parallel processes can produce a wrong output.
Fig. 10 Possible effect of a single event effect corrupting critical or shared resources of a parallel device. As a reult of the impact, multiple parallel processes can produce a wrong output.CC-BY-NC-ND · From the PDF
(a) Expected
(a) ExpectedCC-BY-NC-ND · From the PDF
Image 12
CC-BY-NC-ND · From the PDF
Image 13
CC-BY-NC-ND · From the PDF
(b) Tolerable
(b) TolerableCC-BY-NC-ND · From the PDF
(d) False positive (e) Misclassification
(d) False positive (e) MisclassificationCC-BY-NC-ND · From the PDF
Fig. 12 Fault propagation and reliability evaluation methodologies in complex computing devices. The fault is originated in the physical transistor, then propagates to the circuit, the microarchitecture, and eventually to the software output. Beam experiments evaluate the probability for a fault in the physical layer to be generated and propagate till the software output. Fault injection at different levels of abstractions can be used to better understand fault propagation. Evaluations closer to the physical layer are more realistic, while evaluations closer to the software layer are more efficient.
Fig. 12 Fault propagation and reliability evaluation methodologies in complex computing devices. The fault is originated in the physical transistor, then propagates to the circuit, the microarchitecture, and eventually to the software output. Beam experiments evaluate the probability for a fault in the physical layer to be generated and propagate till the software output. Fault injection at different levels of abstractions can be used to better understand fault propagation. Evaluations closer to the physical layer are more realistic, while evaluations closer to the software layer are more efficient.CC-BY-NC-ND · From the PDF
Image 17
CC-BY-NC-ND · From the PDF
Fig. 14 Description of the basic concept of ABFT applied to matrix multiplication (top) and smart pooling (bottom). These hardening solutions are highly effective in reducing the impact of radiation in CNNs execution [11].
Fig. 14 Description of the basic concept of ABFT applied to matrix multiplication (top) and smart pooling (bottom). These hardening solutions are highly effective in reducing the impact of radiation in CNNs execution [11].CC-BY-NC-ND · From the PDF
Image 19
CC-BY-NC-ND · From the PDF
Fig. 15 Structure of the error detection framework implemented comparing subsequent input images and the correspondent detection. Similar input frames should provide similar detection [168].
Fig. 15 Structure of the error detection framework implemented comparing subsequent input images and the correspondent detection. Similar input frames should provide similar detection [168].CC-BY-NC-ND · From the PDF
Image 21
CC-BY-NC-ND · From the PDF
Fig. 16 Redundant low-precision functional units can be used to implement efficient DWC strategies. On GPUs, when a code is executed in a given precision the functional units of different precision are idle and can be leveraged for duplication [176].
Fig. 16 Redundant low-precision functional units can be used to implement efficient DWC strategies. On GPUs, when a code is executed in a given precision the functional units of different precision are idle and can be leveraged for duplication [176].CC-BY-NC-ND · From the PDF
Fig. 17 Fault-aware training is a promising solution to make the DNN more reliable [179].
Fig. 17 Fault-aware training is a promising solution to make the DNN more reliable [179].CC-BY-NC-ND · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

  • MobileNetV2: Inverted Residuals and Linear Bottlenecks

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00474 · Cited by 28,755 · cited 1×

    “Building upon this function, several other ReLU-like activations have been developed, e.g., the scaled exponential linear unit (SELU) [27], the Gaussian error linear unit (GELU) [28], and ReLU6 [29].”
    Background
  • An introduction to ROC analysis

    Fawcett, T. (2006) · Pattern Recognition Letters · doi:10.1016/j.patrec.2005.10.010 · Cited by 21,879 · cited 2×

    “To evaluate error criticality for detection, it is necessary to consider the BB’s area, adopting the methodology used in the image processing community [61].”
    Method
  • Large-Scale Machine Learning with Stochastic Gradient Descent

    Bottou, L. (2010) · Proceedings of COMPSTAT'2010 · doi:10.1007/978-3-7908-2604-3_16 · Cited by 5,795 · cited 1×

    “Given the gradients, an optimizer, e.g., stochastic gradient descent [23], updates the weights in order to minimize the loss function.”
    Method
  • Basic concepts and taxonomy of dependable and secure computing

    Avizienis, A., Laprie, J.-C., Randell, B., & Landwehr, C. (2004) · IEEE Transactions on Dependable and Secure Computing · doi:10.1109/tdsc.2004.2 · Cited by 5,223 · cited 1×

    “In this work, we will use the taxonomy defined by Avizienis et al. [60].”
    Method
  • Backpropagation through time: what it does and how to do it

    Werbos, P. J. (1990) · Proceedings of the IEEE · doi:10.1109/5.58337 · Cited by 4,879 · cited 1×

    “ANNs are universal function approximators that, thanks to Backpropagation (short for backward propagation of errors) training [20], [21], [22] and sufficient network complexity, enable the solution of a variety of tasks, e.g., classification, detection, and regression.”
    Background
  • Gaussian Error Linear Units (GELUs)

    Hendrycks, D., & Gimpel, K. (2016) · arXiv · doi:10.48550/arxiv.1606.08415 · Cited by 3,537 · cited 1×

    “Building upon this function, several other ReLU-like activations have been developed, e.g., the scaled exponential linear unit (SELU) [27], the Gaussian error linear unit (GELU) [28], and ReLU6 [29].”
    Background
  • Deep Learning using Rectified Linear Units (ReLU)

    Agarap, A. F. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1803.08375 · Cited by 2,399 · cited 1×

    “…most computationally demanding and, thus, vulnerable to radiation), activation functions are used for ensuring a nonlinear input–output relationship in DNNs and are very often implemented through rectified linear units (ReLUs) [25] ReLU ( x ) = max ( 0 , x ), where x stands for the input tensor.”
    Method
  • Self-Normalizing Neural Networks

    Klambauer, G., Unterthiner, T., Mayr, A., & Hochreiter, S. (2017) · arXiv · doi:10.48550/arxiv.1706.02515 · Cited by 2,267 · cited 1×

    “Building upon this function, several other ReLU-like activations have been developed, e.g., the scaled exponential linear unit (SELU) [27], the Gaussian error linear unit (GELU) [28], and ReLU6 [29].”
    Background
  • NVIDIA Tesla: A Unified Graphics and Computing Architecture

    Lindholm, E., Nickolls, J., Oberman, S., & Montrym, J. (2008) · IEEE Micro · doi:10.1109/mm.2008.31 · Cited by 1,362 · cited 1×

    “Each basic computing unit (named CUDA core in NVIDIA devices) in the SM executes one thread with dedicated registers, avoiding complex resource sharing or the need for long pipelines [37] [see Fig.”
    Background
  • Pattern Recognition and Machine Learning

    Maindonald, J. (2007) · Journal of Statistical Software · doi:10.18637/jss.v017.b05 · Cited by 1,349

Show all 165 references

Research works printed but not linked: 16

  1. 2023, NVIDIA Ampere GPU Whitepaper
  2. 2023, Google Coral Edge TPU Explained in Depth
  3. 2018, NVIDIA Jetson AGX Xavier Developer Kit | NVIDIA Developer
  4. 2011, Preview Road Vehicles Functional Safety
  5. Sumit, 2018, A Comprehensive Guide to Convolutional Neural Networks
  6. University, 2023, CS231n Convolutional Neural Networks for Visual Recognition
  7. Bengio, 2017, Deep Learning, 1
  8. 2022, Zynq-7000 All Programmable SoC
  9. 2006, Measurement and Reporting of Alpha Particle and Terrestrial Cosmic Ray-Induced Soft Errors in Semiconductor Devices
  10. Ziegler, 2004, SER—History, Trends and Challenges: A Guide for Designing With Memory ICs
  11. 2020, NVBitFI: An Architecture-level Fault Injection Tool for GPU Application Resilience Evaluations
  12. 2016, ESCC Basic Specification No. 22900
  13. 2016, Carol Radiation Setup
  14. Wyrwas, 2022, Proton Testing of AMD v1202b System on Chip
  15. Wyrwas, 2022, Radiation Effects and Analysis Lessons: A Scientist’s Field Instruction to Explain Radiation Testing
  16. Navarro, 2021, High-Performance Compute Board—A Fault-Tolerant Module for On-Boards Vision Processing

Cited by

See all 97

Metadata

Authors
Rech, P.
Affiliations
University of Trento (IT)
Corresponding author
Rech, Paolo
Venue
IEEE Transactions on Nuclear Science
Published
2024-04
Volume
71
Issue
4
Pages
377-404
ISSN
  • 0018-9499 (ISSN-L)
  • 1558-1578
OpenAlex
W4390576801
Cited by
75
References
165
Citations by year
2026: 33 · 2025: 29 · 2024: 12
FWCI
7.65
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Nuclear Science (publishedVersion, cc-by-nc-nd)
  • Institutional Research Information System (Università degli Studi di Trento) (submittedVersion)
  • ePubs (Science and Technology Facilities Council, Research Councils UK) (submittedVersion, cc-by-nc-nd)
Funders
  • U.S. Department of Energy
  • Università degli Studi di Trento
  • Universidade Federal do Rio Grande do Sul
  • Los Alamos National Laboratory
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Radiation Effects in Electronics
  • Physical Sciences › Computer Science › Software › Software Reliability and Analysis Research
  • Physical Sciences › Computer Science › Hardware and Architecture › VLSI and Analog Circuit Testing
Keywords
artificial neural networks, deep neural network, reliability, safety-critical systems, space applications, radiation effects, fault propagation, fault tolerance, hardware reliability, fault sensitivity analysis, autonomous vehicles, aerospace

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Artificial Neural Networks for Space and Safety-Critical Applications: Reliability Issues and Potential Solutions. https://app.cochat.ai/papers/10.1109/tns.2024.3349956
The paper
Rech, P. (2024). Artificial Neural Networks for Space and Safety-Critical Applications: Reliability Issues and Potential Solutions. IEEE Transactions on Nuclear Science. https://doi.org/10.1109/tns.2024.3349956

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.