Behavioral Fault Model for Neural Networks

The term neural network (NN) originally referred to a network of interconnected neurons which are basic building blocks of the nervous system. Fault tolerance is known as an inherent feature of artificial neural networks (ANNs). Wide attention has been given to the problem of fault-tolerance in VLSI implementation domain and not enough attention has been paid to intrinsic capacity of survival to faults. In this work we focus on the impact of faults on the neural computation in order to show neural paradigms cannot be considered intrinsically fault-tolerant. A high abstraction level (corresponding to the neural graph) error model is introduced in this paper. We propose fault model and present an analysis of the usability of our method for fault masking. Simulation results show with this new fault model, the fault with less significant contribution is masked in output.