Learnings and applications of feedforward nets
L.K. Li · 2003
A three-layer network which approximates a desired function f by a piecewise constant function is constructed. Backpropagation and classic gradient learning are present. A learning method is presented which gives the optimal weights at each iteration. Applications to pattern recognition are given with discussions on using RBF (radical basis function) unit networks. In addition, it is proved that the error is bounded by a linear function of the grid size.>