SEU susceptibility analysis of a feedforward neural network implemented in a SRAM-based FPGA
Israel C. Lopes, Fernanda Lima Kastensmidt, Altamiro Susin · 2017
Artificial Neural Networks (ANNs) have gained a considerable interest in clustering, pattern recognition, function approximation and many others applications, due to its parallel capability of processing the data. Moreover, ANNs can also be used to accelerate parts out of an algorithm which the data can be approximated using NPUs (Neural Processing Units). FPGAs are intrinsically parallel making these devices a good choice to implement an ANN. However, SRAM-based FPGAs are very susceptible to neutron-induced soft errors. Consequently, an ANN running in SRAM-based FPGAs devices must be characterized under soft errors. This work shows the soft error effects of a feedforward ANN under fault injection emulation performed in the bitstream of the FPGA. The number of critical bits able to provoke errors and failures in the output of the network are analyzed. Results showed the reliability superiority of the proposed ANN over the conventional implementation of the same image processing performed by the ANN.