Ellipsoidal Function Modulated ART Neural Networks for Pattern Recognition

In this paper, we propose a neural network that adopts the structure of the instar-outstar pair of the ART neural networks, uses the equivalent Gaussian functions of the training pattern clusters to substitute the weight vectors of the in star blocks, and two receptive fields of the covariance matrices of the equivalent Gaussian functions of the training pattern clusters to form the hyper-ellipsoidal contours to substitute the weight vectors of the out star blocks. And we call this the Ellipsoidal Function Modulated ART (EFM-ART) neural network. The proposed neural network adopts the fundamental structure of the ART neural networks but omits many other genius functions of the ART neural networks. For observation and feasibility evaluation, the vectors of the intensity of the wavelet packet parameters of sounds are adopted as the vectors of feature parameters. Simulation results can highly support the feasibility of this EFM-ART for pattern recognition and the capability in handling the problem of stability and plasticity dilemma as done by many other ART neural networks.