Through Metric learning techniques, a metric function is learned, which shows how similar/dissimilar two samples are. From the perspective of feature selection, metric learning can be represented as a transform function mapping each sample into a new point in the new feature space. Geometric Mean Metric Learning (GMML) is one of promising methods which achieve good performance in terms of accuracy and time complexity. In this paper, we propose the use of GMML algorithm in a neural network to perform Riemannian computing on the SPD matrices which improves accuracy and reduces time complexity. We also use the eigenvalue rectification layer as a non-linear activation function to enhance the non-linearity of our model. Experimental evaluations on several benchmark data sets demonstrate that the proposed method improves accuracy in comparison with the state-of-the-art approaches.

Keywords