Selective attention is modeled by Support Vector Machine to recognize Super-Imposed patterns. Selective attention makes human to recognize desired information against complex background noises and gather imposed patterns. and SVM is the newest method for classification. But, SVM deals data not on the original data space but on the transformed kernel space. In this case, transformed kernel space has much higher or infinite dimension than original data space. If we select proper kernel function and then on kernel space transformed data distribution is well formed and linearly separable-like, SVM can provide much better classification accuracy then on original data space. To combine the mechanism of selective attention in engineering sense, gradient ascent adaptation algorithm is used. With the slight extention of gradient ascent algorithm, input pattern can be modified to the more favorable-to-recognize one and this filtered input pattern comes to be the one which is memorized in the network during training phase.
Withrawing in the right place at right moment, is very critical in the selective attention method. The boundary of Support Vector can be used to stopping the attention processing as criterion. Following things can be used to measure the value of class; super-imposed input patterns, the distance between filtered pattern and the source, the SVM output after the attention, hidden neuron's maximum value which has closest value towards initial status. When certain input pattern is correct, the distance will be relatively small, consquently the SVM output and hidden neuron's maximum will be relatively large. Proposed stop-point criterion and confidence measure were tested in Super-Imposed handwritten digits which consisted of USPS DataSet. Recognition rate got much improved when selective attention was used instead of SVM only.
중첩패턴을 인식하기 위해 SVM을 기반으로한 선택적 주의집중 모델을 공학적으로 구현하였다. 강력한 학습이론인 SVM과 잡음환경에서 강한 인간의 선택적 주의집중을 접목시켜 단일 인식기로는 인식이 불가능한 중첩패턴에 대해 인식을 시도하였고, 그 결과 인식율의 높은 향상이 있었다. 선택적 주의집중을 접목하기 위해 gradient ascent 알고리즘을 사용했고, 정지점 척도를 위해 지지벡터의 경계를 기준으로 삼았다. 일반적인 주의집중을 그대로 사용하면 모순이 존재하기 때문에 변형된 입력이 원래 입력의 범위를 넘지 못하는 제한을 두었고, 실험은 USPS 데이타를 통해 이루어졌으며 단일 인식기를 사용했을 때보다 훨씬 우수한 인식시스템임을 증명했다.
