Analyzing the reliability of convolutional neural networks on GPUs : GoogLeNet as a case study
Convolutional Neural Networks (CNNs) are used for tasks such as object recognition. Once a CNN model is used in a radiative environment, reliability of the system against soft errors is a crucial issue, especially in safety-critical and high-performance applications that bound with real-time respons...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English |
| Published: |
Institute of Electrical and Electronics Engineers Inc.
2020
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/42444/1/Analyzing%20the%20reliability%20of%20convolutional%20neural%20networks%20on%20GPUs.pdf http://umpir.ump.edu.my/id/eprint/42444/2/Analyzing%20the%20reliability%20of%20convolutional%20neural%20networks%20on%20GPUs_GoogLeNet%20as%20a%20case%20study_ABS.pdf http://umpir.ump.edu.my/id/eprint/42444/ https://doi.org/10.1109/ICCIT-144147971.2020.9213804 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Convolutional Neural Networks (CNNs) are used for tasks such as object recognition. Once a CNN model is used in a radiative environment, reliability of the system against soft errors is a crucial issue, especially in safety-critical and high-performance applications that bound with real-time response. Selectively-hardening techniques do improve the reliability of these systems. However, the hard question in selective techniques is "how to exclusively select code portions to harden, to safe the performance from being degraded". In this paper, we propose a comprehensive analysis methodology for CNN-based classification models to confidently determine the only vulnerable parts of the source code. To achieve this, we propose a technique, Layer Vulnerability Factor (LVF) and adopt another technique, Kernel Vulnerability Factor (KVF). We apply these techniques to GoogLeNet, which is a famous image classification model, to validate our methodology. We precisely identify the parts of the GoogLeNet model that need to be hardened instead of using expensive duplication solutions. |
|---|