Practical Neural Network Recipies in C++

196

Chapter 11

input to an intermediate level, it is first clamped to one extreme, then

the other. If the output neuron corresponding to a certain class tends

to become more activated (across the training set) when the input is

clamped fully activated, then it appears that activation of that input

is a feature important to classification into that class. Due to the

unpredictable nature of interactions among inputs, this is a dangerous

technique and is easily abused.

Another version of sensitivity analysis presents a case to the

network, then observes the effect of varying individual inputs while all

others are fixed. Changing some inputs ma ...