Discuss the neuron network paper by Cheng and Titterington(1994, Statistical Science 1994, page 2-30)
1. The McCulloch-Pitts neuron : binary input and output, connection weight,bias
2. Training method: compared with least squares ; steepest decent.
(note : 1. NN trains sample one at a time.
2. The least squares leads to Fisher's linear discriminant rule)
3. Generalization : 3.1. other activation functions (has to be monotonic)
3.2 . hidden layers.
y= max (x_1, x_2)
(need 4 nodes in the hidden layer)
(Note: it is not uncommon to have more hidden units than input units; anti-dimension
reduction?)
3.3. Error backpropagation method.