next up previous contents index
Next: Generate Inputs Sets Up: Analysis of Trained Network Previous: Representing the Input Space

   
Observing Hidden Unit Activations

Another way of interpreting the hidden layer functionality is to observe the density of activations in each hidden unit. This observation can be done on the output of the hidden unit, i.e. the result of the activation function, or at an earlier stage, i.e. the net input of the hidden unit. The density plot will give the density distribution of possible activations of the hidden unit, given the set of input patterns.

There are two ways of displaying these density plots (see Section 4.2.2) :

Jittered Density Plots[Berkeley95] ,
where high number of input patterns resulting a certain output range is observed by dense scattered points in a plot seen in Figure 5 (use tool described in Section 4.2.2) and,
  
Figure 5: A jittered density plot.

\resizebox*{0.9\textwidth}{!}{\includegraphics{jitter_3.eps}}




Histogram Plots ,
where the number of input patterns resulting in the activations are accumulated in bins along the output range in plot as seen in Figure 6 (use tool described in Section 4.2.2).
  
Figure 6: The histogram plot for the same unit for which the jittered density plot was given in Figure 5.

\resizebox*{0.9\textwidth}{!}{\includegraphics{histo_2.eps}}



next up previous contents index
Next: Generate Inputs Sets Up: Analysis of Trained Network Previous: Representing the Input Space
Cengiz Gunay
2000-06-25