On the Reliability of Convolutional Neural Network Implementation on SRAM-based FPGA
Authors:
- 1 October 2019
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Abstract
In recent years, topics around machine learning and artificial intelligence (AI) have (re-)gained a lot of interest due to high demand in industrial automation applications in various areas such as medical, automotive and space and the increasing computational power offered by technology advancements. One common task for these applications is object recognition/classification whose input is usually an image taken from camera and output is whether an object is present and the class of the object. In industrial pipeline, this task could be used to identify possible defects in products; in automotive application, such task could be deployed to detect pedestrians for Advanced Driver-Assistance Systems (ADAS). When the task is safety-critical as in automotive application, the reliability of the task implementation is crucial and has to be evaluated before final deployment. On the other hand, Field Programmable Gate Array (FPGA) devices are gaining increasing attention in the hardware acceleration part for machine learning applications due to their high flexibility and increasing computational power. When the SRAM-based FPGA is considered, Single Event Upset (SEU) in configuration memory induced by radiation particle is one of the major concerns even at sea level. In this paper, we present the fault injection results on a Convolutional Neural Network (CNN) implementation on Xilinx SRAM-based FPGA which demonstrate that though there exists built-in redundancy in CNN implementation one SEU in configuration memory can still impact the task execution results while the possibility of Single Event Multiple Upsets (SEMU) must also be taken into consideration.Keywords
This publication has 15 references indexed in Scilit:
- Real-time grasp detection using convolutional neural networks2015
- Deep learning in neural networks: An overviewNeural Networks, 2015
- Fixed-point feedforward deep neural network design using weights +1, 0, and −12014
- Large-Scale Video Classification with Convolutional Neural Networks2014
- Convolutional Neural Networks for Sentence Classification2014
- Neutron radiation test of graphic processing units2012
- Effects of atmospheric neutrons on devices, at sea level and in avionics embedded systemsMicroelectronics Reliability, 2007
- Investigating the Fault Tolerance of Neural NetworksNeural Computation, 2005
- Gradient-based learning applied to document recognitionProceedings of the IEEE, 1998
- Single event upset at ground levelIEEE Transactions on Nuclear Science, 1996