In this paper we present new algorithms for training reduced-size nonlinear representations in the Kernel Dictionary Learning (KDL) problem. Standard KDL has the drawback of a large size of the kernel matrix when the data set is large. There are several ways of reducing the kernel size, notably Nyström sampling. We propose here a method more in the spirit of dictionary learning, where the kernel vectors are obtained with a trained sparse representation of the input signals. Moreover, we optimize directly the kernel vectors in the KDL process, using gradient descent steps. We show with three data sets that our algorithms are able to provide better representations, despite using a small number of kernel vectors, and also decrease the execution time with respect to KDL.