Introduction to Neural Networks for C#, Session 9

Session Material

In pervious class session you've seen that a feedforward neural network can have zero or more hidden layers, and some number of neurons in each of these hidden layers. How do you know how many hidden layers to have, or how many neurons to have in each of those hidden layers. Pruning is one tool that you have to help you decide upon a structure for your neural network. Pruning only addresses the hidden layers. How you structure your input and output layer define the very nature of the problem that you are trying to solve.

By using pruning you can create a more optimal hidden layer structure. There are two types of pruning that will be discussed in this chapter.

• Incremental Pruning
• Selective Pruning

Incremental Pruning starts with just input and output layers. The algorithm will incrementally increase the size of the neural network and retrain at each increment. Whichever configuration trained the best is assumed to be the optimal network configuration.

Selective Pruning starts with an already trained neural network that has some hidden layers and neurons. The selective pruning algorithm will choose hidden neurons to remove that will not affect the error rate of the neural network. Through this process unproductive neurons can be removed.

Videos for this Session

Video	Title
	Introduction to Neural Networks for C#(Class 9/16, Part 1/3)
	Introduction to Neural Networks for C#(Class 9/16, Part 2/3)
	Introduction to Neural Networks for C#(Class 9/16, Part 3/3)