Feature Extraction by Measuring Importance of Inputs

Various methods have been proposed to measure the importance of inputs (Sarle, 1998) and are likely to be useful in different applications of neural nets. The two most common notions of importance are predictive importance and causal importance. Predictive importance is concerned with the increase in generalization error when an input is omitted from a network. Causal importance is concerned with situations in which an individual wants to quantify the relationship between input value manipulation and consequent output change. 

In linear models, the weights have a simple interpretation each weight is the change in the output associated with a unit change in the corresponding input, assuming all other 

---