On the evaluation of relevance learning by a multi-layer perceptron

In this paper, we introduce a novel method of relevance learning by a multi-layer perceptron. The relevance learning is regarded as learning from the relationship among two or more outputs of the network. The learning network architecture is based on a simple multi-layer perceptron with a modified back-propagation learning algorithm. Unlike the conventional multi-layer perceptron that learns from a set of an input feature vector and the target output, the proposed network can obtain a nonlinear mapping between a set of two or more vector inputs and the desired relevance. For instance, the desired relevance represents the dissimilarity among given objects. We show the performance of the proposed network with some experiments with four artificially generated data set. We then discuss the theoretical and mathematical background underlying the network learning with some related works. We evaluate the obtained arrangement of objects in comparison with the result of principle component analysis (PCA) and multidimensional scaling method (MDS). This work also contributes to the measurement of human subjective evaluation for multidimensional perceptual scaling. Some experimental results on the low-dimensional representation of color hue data set and emotional facial images were presented.