Assessing the ANN reliability: issues, strategies and opportunities

Ernesto Sanchez Sanchez  - Politecnico di Torino, IT

Abstract :

Nowadays, the computational capabilities reached by solutions based on Artificial Neural Networks are surprisingly overtaking traditional approaches. For example, in 2015 for the very first time, an artificial neural network outperformed a human being in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), an image classification contest. Since that time, and even before, ANNs are ever increasing their presence in safety-critical applications, and this is asking the attention of academy and industrial researches to better understanding how to assess the NN’s resiliency considering hardware faults. In this talk, a description of the main problems related to the reliability of Artificial Neural Networks, as well as possible solutions and research opportunities are presented. The tutorial includes some interesting examples that illustrate the provided concepts.

Sylabus :

The lecture will cover the following topics:

•    Introduction
•    ANN evolution and criticality 
•    Assessment process
•    Fault Mitigation Strategies
•    Reliability assessment examples
•    Conclusions and Future Directions