In this paper we extend the classical neuron by using the kernels to define its corresponding kernel neuron, and construct a training algorithm based on gradient descent scheme for it. For nonlinear classification and regression problems, it can implement the similar performance of feedforward neural networks, e.g. multi-layer perceptrons and radial basis function neural networks. Moreover its algorithm structure is very simple. Experiments further show that it works well on both linear and nonlinear cases.
展开▼
