A Study of Local Training Algorithm for Feedforward Neural Networks andd Face Image Recognition

A Study of Local Training Algorithm for Feedforward Neural Networks andd Face Image Recognition

2005 Masahiro Yoshida

This thesis describes the novel local training algorithms for training feedforward neural networks and the face image recognition by using feedforward nbural networks. After the introduction, in chapter 2, a novel local training based on the RlS(Recursive Least Squares) with penalty function is described. In local training, it is possible to partition the given problem to the neuron level. Therefore, the local trainings improve computational complexity and storage requirements compared to the global approaches. In addition, the penalty function proposed here is derived from a second-order term in the Taylor

expansion of the network output for decreasing the influence of the linearization elror in the RLS. As

result, it is shown that the proposed training algorithm yields the high perfonnance for the convergence rate. In chapter 3, the local training

on the RLS with weight decay is introduced. The proposed algorithm is

on modifying the local training algorithm in chapter 2. As

result,

the equations for weights update among the proposed method is simple, the learning

is faster than the other algorithms. In chapter 4, the face image recognition by using feedforward neural networks has been done. In the previous researches of face image recognition by neural networks, the gray levels on each pixel of the face image have been used for input data. However,

the face image

usually too many pixels,

variety of approaches have

required to reduce the number of the input data. The orthogonal transformations

for reduction of input data and the recognition is

by feedforward neural networks. In chapter 5, the conclusions of this thesis

described.

‑178‑