Construction of Input Layer

Deciding how to construct the input layer is application dependent. It is affected by many considerations. Should fixed-length sequence windows or variable-length sequences be used Is there a dependence on positional information Is it intended to search for signal or search for content What is the importance of local information or global information  

The incorporation of different input features may also come in the form of different neural network designs. One example is cascading networks (Chapter 8), where output from one network becomes the input to the next network. 

An excessive number of input units not only results in poor generalization, but also slows network training. Reducing the number of input variables often leads to improved performance of a given data set, even though information is being discarded. Thus, one important role of pre-processing is to reduce the dimensionality of the input data. 

---