Page 254 -

242 5 Neural Networks

5.25 Design an SVM for classification of the Rocks data into two classes: granites vs.
limestones+marbles. Use features SiO2, CaO and determine experimentally the kernel
with best generalization.

5.26 Design a Kohonen network for the CorkStoppers dataset. Compare the solution
obtained with the supervised classification and with the cluster solution from Exercise
3.7.

5.27 Consider the hierarchical classification of the CTG data, shown in Figure 5.57.
Estimate the bounds for proper learning of the respective MLP6:7:1 and MLP9:5:3
used in the hierarchy.

5.28 Design a majority vote ensemble of MLPs for classifying the Rocks data into the
following classes: granite, diorite, slate, marble and limestone.

5.29 Determine useful predictor variables of the pathologic+suspect classes of the CTG
dataset, using probabilistic neural nets with a genetic algorithm. With these predictors
derive MLP, RBF and SVM solutions for the CTG classification task normal vs.
abnormal. Compare the solutions and assess their generalization capability.

5.30 A Hopfield net is applied for the retrieval of binary images using an array of 10x10
neurons. Five classes of simple geometric shapes are used and the network must
retrieve the prototype that best matches an input image. When applying the network, it
was found that it produced much better results when the images occupied the whole
10x10 array than when they occupied only half of it. Explain why.

5.31 Consider the binary images shown in Figure 5.52. Use the Hopjeld program in the
random serial and full parallel mode with noise corrupted versions of the prototypes
and explain the results obtained, namely for the two-state oscillations in the full parallel
mode.

5.32 Consider the eight digit images shown in Figure 5.55, used as prototypes in a Hopfield
network.
a) Explain how the spurious state shown in Figure 5.56 is formed.
b) Perform experiments of prototype retrieval using noise corrupted images of the
several digits, and determine which pattern is found more often when an incorrecl
retrieval is made. Explain why.