Biometric Authentication: A Machine Learning Approach

5.5. Two-Stage Training Algorithms

The convergence behavior of the BP algorithm for multi-layer networks has always been an issue of great concern. This is because the algorithm tends to be sensitive to initial conditions. Therefore, the convergence problem can be greatly alleviated if proper initial conditions can be estimated. This motivates the two-stage training strategy discussed next.

5.5.1. Two-Stage Training for RBF Networks

The two-stage training of RBF networks starts with the estimation of the kernel means using the K-means algorithm

Equation 5.5.1

Equation 5.5.2

where

Equation 5.5.3

N_j is the number of patterns in the cluster ...