Özet:

Random Fourier features (RFF) provide one of the most prominent means for nonlinear classification in especially large scale data settings. However, considering the original proposal of RFF, Fourier features are randomly drawn from a certain distribution and used unoptimized. In this paper, we investigate Fourier features via a single hidden layer feedforward neural network (SLFN) and optimize, i.e., learn, those features (instead of drawing randomly). The learned Fourier features are deduced from the radial basis function (rbf kernel), and implemented in the hidden layer of the SLFN which is followed by the output layer. We present extensive experiments with 10 different classification datasets from various fields, e.g., bioinformatics. The learning of Fourier features is observed to be highly superior over the competing techniques such as perceptron in the rbf kernel space or a greedy forward feature selection strategy. On the other hand, the Fourier feature learning performs comparably with SVM (support vector machines with rbf kernel) while providing substantial computational benefits, and this is even without using the max margin regularization. Moreover, when tested in wireless mesh networks, the SLFN delivers promising smart steering capabilities.

Anahtar Kelimeler: