1. The distance neuron
• hyperspherical classifiers
• spherical (radial) vs. ellipsoidal
• hard vs. soft neurons
• Gaussian activation functions
2. Unsupervised training of a distance neuron
• simple k-means clustering (LGB)
• ISODATA
• EM algorithms
3. The RBF network
• Architecture
• feedforward with a single hidden layer
• soft distance neurons in hidden layer
• linear perceptrons in output layer
• Universal approximation
• Comparison of MLP and RBF
• local vs. global decisions
• misdetection vs. false alarm
4. Training the RBF network
• Backprop training
• 2-step training
5. The Best-of-all-possible-worlds network
• Augmented perceptrons
• Classification regions
• Use in MLPs
6. The RCE network
• Hard distance neurons
• Training the RCE network

Assignment

1. Perform unsupervised k-means clustering for the Iris or Peterson Barney data (whichever you used for the BP assignment).
2. Using the results of the previous exercise train the second layer of an RBF using the PLA (coded in assignment 3).
3. Compare the generalization of the RBF with that of the MLP.