Similarity measurement in convolutional space
Hosseinali Ghiassirad, Mohammad Teshnehlab · 2012
We introduce a method to find the difference of data based on convolution. The method may be used to recognize or verify training data where the number of data classes is unknown or very large in training phase or used as a kernel in an SVM or in an RBF Network. A common solution is to train a machine to maps two input patterns into a new space, which may be high dimensional, such that the output value of the machine approximates the “semantic” distance of input pair. The learning algorithm minimizes an error function which measures the similarity of pair of patterns presented on input layer. In the best case the error function is zero for genuine pairs and is infinity for impostor pairs. The similarity measure is done in convolutional space with Convolutional Neural Network which is robust to spatial distortions. The method is applied on AT&T dataset for face verification task.