Abstract:In this paper, firstly, the effect of stuck at faults and error inputs on single neural network node is analysed in detail. Next, the behavior of feedforward neural networks with a hard-limit activation function under stuck-at faults is studied, and the relationship between the network performance and the number of faulty interconnections, together with the method for analyzing the faulty behavior, is presented. The Monte Carlo simulations show that the method proposed in this paper is correct. Lastly, the analysis results indicate that the smaller the scale of such networks is, the more fault-tolerant they will be.

Faulty behavior of feedforward neural networks