1.Discuss error backpropagation
2. Discuss LeCun et al.
    2.1.   Number of parameters  estimated > number of  cases.
    2.2.   Details of estimation do matter.
    2.3.   When to stop is also important.
    2.4.    Network design.

3. Statistical connection.
    (When the number of connection parameters is relatively small,
       compared to the sample size)

   3.1 Logistic regression
   3.2. Nonlinear regression,   stochastic approximation (learning rate problem),
           asymptotics for single hidden-layer  netweok
          (White, 1989, JASA, page 1003-1013)
   3.3 Projection pursuit regression.
   3.4.  Dimension reduction model.
   3.5. Brillinger's result.