Metrics for Weight Stuck-at-Zero Fault in Sigmoidal FFANNs

In this paper, a class of weight fault model known as single weight stuck-at-zero for a single hidden layer (with sigmoidal nodes) feed forward artificial neural networks is analyzed. Fault measures/metrics are derived for weight stuck-at-zero fault. Experiments are conducted for four function approximation tasks wherein a set of 30 networks are trained for each task. A network which has a least validation error is selected for further analysis of weight stuck-at fault for the four function approximation tasks. The average change in the prediction error on single fault seeding is measured and compared with the predicted fault measure. Correlation between the derived fault measures and empirical measure of single fault seeding demonstrate that the correlation is significant at 0.10 level for both the derived measures, for one o f the derived measures, the correlation is significant at 0.05 level. Thus, these two derived measures are shown to be good metrics for the measurement of the fault tolerance of the network to single weight fault, at least for the two function approximation tasks. Further experimentation is required to empirically assess the validity of these measures.