Assessment of Machine Learning Algorithms for Near-Sensor Computing Under Radiation Soft Errors

International audienceMachine learning (ML) algorithms have been regaining momentum thanks to their ability to analyze substantial and complex data, supporting artificial intelligence decisions in cloud computing but also in near-sensor computing in endpoint devices. Both cloud and near-sensor computing are liable to radiationinduced soft errors, especially in automotive and aerospace safety-critical applications. In this regard, this paper contributes by comparing the accuracy of tw o prominent machine learning algorithms running on a lowpow er processor upset by radiation-induced soft errors. Both ML algorithms have been assessed w ith the help of a fault injectionbased method able to natively emulate soft errors directly in a development board. In addition, neutron radiation test results suggest the most critical situations in w hich mitigation solutions should address