Implementation of Robust Feedforward Neural Network Using Classifier Structure

수렴성 구조를 이용한 강인한 선행 신경망 구현

In this paper, we improve feedforward neural network performance by eliminating the effect of gross error using classifier structure. At first, we prove the output of classifier converges to the posteriori probability of each pattern given input x, $f_0({\theta}_1|x)$. And we apply filtering approach based on the robust statistics before reconstructing continuous output. The data distorted with noise can be rejected by this process. Finally, we suggest neurofilter structure. Simulation result shows that our structure yields consistent estimates even in the presence of noise.