Details
| Original language | English |
|---|---|
| Title of host publication | Computer Safety, Reliability, and Security |
| Subtitle of host publication | 42nd International Conference, SAFECOMP 2023, Toulouse, France, September 20–22, 2023, Proceedings |
| Editors | Jérémie Guiochet, Stefano Tonetta, Friedemann Bitsch |
| Publisher | Springer International Publishing AG |
| Pages | 243–256 |
| Number of pages | 14 |
| ISBN (electronic) | 978-3-031-40923-3 |
| ISBN (print) | 978-3-031-40922-6 |
| Publication status | Published - 2023 |
Publication series
| Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
|---|---|
| Volume | 14181 LNCS |
| ISSN (Print) | 0302-9743 |
| ISSN (electronic) | 1611-3349 |
Abstract
Neural networks (NNs) are commonly used for environmental perception in autonomous driving applications. Safety aspects in such systems play a crucial role along with performance and efficiency. Since NNs exhibit enormous computational demands, safety measures that rely on traditional spatial or temporal redundancy for mitigating hardware (HW) faults are far from ideal. In this paper, we combine algorithmic properties with dedicated HW features to achieve lightweight fault tolerance. We leverage that many NNs maintain their accuracy when quantized to lower bit widths and adapt their quantization configuration during runtime to counteract HW faults. Instead of masking computations that are performed on faulty HW, we introduce a fail-degraded operating mode. In this mode, reduced precision computations are exploited for NN operations, as opposed to fully losing compute capability. This allows us to maintain important synapses of the network and thus preserve its accuracy. The required HW overhead for our method is minimal because we reuse existing HW features that were originally implemented for functional reasons. To demonstrate the effectiveness of our method, we simulate permanent HW faults in a NN accelerator and evaluate the impact on a NN’s classification performance. We can preserve a NN’s accuracy even at higher error rates, whereas without our method it completely loses its prediction capabilities. Accuracy drops in our experiments range from a few percent to a maximum of 10 %, confirming the improved fault tolerance of the system.
Keywords
- Approximate Computing, Automotive, Fault Tolerance, Neural Network Hardware, Neural Networks, Quantization
ASJC Scopus subject areas
- Mathematics(all)
- Theoretical Computer Science
- Computer Science(all)
- General Computer Science
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
Computer Safety, Reliability, and Security: 42nd International Conference, SAFECOMP 2023, Toulouse, France, September 20–22, 2023, Proceedings. ed. / Jérémie Guiochet; Stefano Tonetta; Friedemann Bitsch. Springer International Publishing AG, 2023. p. 243–256 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 14181 LNCS).
Research output: Chapter in book/report/conference proceeding › Contribution to book/anthology › Research › peer review
}
TY - CHAP
T1 - Online Quantization Adaptation for Fault-Tolerant Neural Network Inference
AU - Beyer, Michael
AU - Borrmann, Jan Micha
AU - Guntoro, Andre
AU - Blume, Holger
N1 - This work is supported by the German federal ministry of education and research (BMBF), project ZuSE-KI-AVF (grant no. 16ME0062).
PY - 2023
Y1 - 2023
N2 - Neural networks (NNs) are commonly used for environmental perception in autonomous driving applications. Safety aspects in such systems play a crucial role along with performance and efficiency. Since NNs exhibit enormous computational demands, safety measures that rely on traditional spatial or temporal redundancy for mitigating hardware (HW) faults are far from ideal. In this paper, we combine algorithmic properties with dedicated HW features to achieve lightweight fault tolerance. We leverage that many NNs maintain their accuracy when quantized to lower bit widths and adapt their quantization configuration during runtime to counteract HW faults. Instead of masking computations that are performed on faulty HW, we introduce a fail-degraded operating mode. In this mode, reduced precision computations are exploited for NN operations, as opposed to fully losing compute capability. This allows us to maintain important synapses of the network and thus preserve its accuracy. The required HW overhead for our method is minimal because we reuse existing HW features that were originally implemented for functional reasons. To demonstrate the effectiveness of our method, we simulate permanent HW faults in a NN accelerator and evaluate the impact on a NN’s classification performance. We can preserve a NN’s accuracy even at higher error rates, whereas without our method it completely loses its prediction capabilities. Accuracy drops in our experiments range from a few percent to a maximum of 10 %, confirming the improved fault tolerance of the system.
AB - Neural networks (NNs) are commonly used for environmental perception in autonomous driving applications. Safety aspects in such systems play a crucial role along with performance and efficiency. Since NNs exhibit enormous computational demands, safety measures that rely on traditional spatial or temporal redundancy for mitigating hardware (HW) faults are far from ideal. In this paper, we combine algorithmic properties with dedicated HW features to achieve lightweight fault tolerance. We leverage that many NNs maintain their accuracy when quantized to lower bit widths and adapt their quantization configuration during runtime to counteract HW faults. Instead of masking computations that are performed on faulty HW, we introduce a fail-degraded operating mode. In this mode, reduced precision computations are exploited for NN operations, as opposed to fully losing compute capability. This allows us to maintain important synapses of the network and thus preserve its accuracy. The required HW overhead for our method is minimal because we reuse existing HW features that were originally implemented for functional reasons. To demonstrate the effectiveness of our method, we simulate permanent HW faults in a NN accelerator and evaluate the impact on a NN’s classification performance. We can preserve a NN’s accuracy even at higher error rates, whereas without our method it completely loses its prediction capabilities. Accuracy drops in our experiments range from a few percent to a maximum of 10 %, confirming the improved fault tolerance of the system.
KW - Approximate Computing
KW - Automotive
KW - Fault Tolerance
KW - Neural Network Hardware
KW - Neural Networks
KW - Quantization
UR - http://www.scopus.com/inward/record.url?scp=85172099424&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-40923-3_18
DO - 10.1007/978-3-031-40923-3_18
M3 - Contribution to book/anthology
SN - 978-3-031-40922-6
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 243
EP - 256
BT - Computer Safety, Reliability, and Security
A2 - Guiochet, Jérémie
A2 - Tonetta, Stefano
A2 - Bitsch, Friedemann
PB - Springer International Publishing AG
ER -