Kernel CMAC: an Efficient Neural Network for Classification and Regression
Gábor Horváth · 2006
Abstract: Kernel methods in learning machines have been developed in the last decade as new techniques for solving classification and regression problems. Kernel methods have many advantageous properties regarding their learning and generalization capabilities, but for getting the solution usually the computationally complex quadratic programming is required. To reduce computational complexity a lot of different versions have been developed. These versions apply different kernel functions, utilize the training data in different ways or apply different criterion functions. This paper deals with a special kernel network, which is based on the CMAC neural network. Cerebellar Model Articulation Controller (CMAC) has some attractive features: fast learning capability and the possibility of efficient digital hardware implementation. Besides these attractive features the modelling and generalization capabilities of a CMAC may be rather limited. The paper shows that kernel CMAC – an extended version of the classical CMAC network implemented in a kernel form – improves that properties of the classical version significantly. Both the modelling and the generalization capabilities are improved while the limited computational complexity is maintained. The paper shows the architecture of this network and presents the relation between the classical CMAC and the kernel networks. The operation of the proposed architecture is illustrated using some common benchmark problems. Keywords: kernel networks, input-output system modelling, neural networks, CMAC, generalization error