A multi-class classification technique which combines kernel logistic neural network (KLNN) and restricted Boltzmann machine (RBM), called KLNN-RBM, is designed. The principal component analysis (PCA) is applied to determine the dimension of the kernel function. The initial weights and thresholds of this model are obtained by RBM. Then, the maximum likelihood estimate with a ridge regularization term and a new stochastic gradient descent method with a scaling factor are used to optimize the parameters in order to realize the multi-class classification. Some numerical simulations illustrate the validity of the proposed method.