Evaluating neural network robustness with an architecture built around L-Neuro 2.3

International audienceWe study the sensitivity of an Artificial Neural Network designed to classify textures in satellite images, with respect to a particular kind of fault, so-called Single Event Upset. These faults are likely to occur as a consequence of interaction with radiations (space, nuclear) and result, for digital microcircuits, in a transient modification (bit flip) of memorized bits of information. Results of fault simulations performed on a digital implementation using a neural architecture built around the L-Neuro2.3 chip from PhilipsÝ are presented. Particularly, we study the impact on the network classification performances of errors in the bits of the input stimuli and synaptic weights, as well as on the memory storing the program emulating the neural network itself