Practical Neural Network Recipies in C++

250

Chapter 15

[Davis and Hwang, 1992] take a different route toward the

same end. First they fully train the network using the entire training

set. Then they determine which members of the training set are

closest to decision boundaries. These potentially troublesome samples

are used to continue training the network. They report that this two-

tiered approach can significantly improve performance.

Hidden Bias

One of the most dangerous phenomena in training set construction is

accidental incorporation of human biases. These are learned by the

network just as easily as any other pattern. The problem is that we

may not know that the bias was learne ...