Extended Sdm Model And Its Application In The Prediction And Identification

Kanerva's Sparse Distributed Memory (SDM) can realize the problem of training and recognizing patterns with large dimension due to its sparse address choice and data storage in the distributed mode. It simulates the partial function of human cerebellum to a certain extent, generalizes random access mode of current computers and is applied in many domains such as pattern recognition and associative memories. Because of randomness of presetting the address matrix between the input and hidden layers, SDM produces unfavorable consequences for classification when handling unevenly distributed patterns. And its learning mode of outer product results in poor nonlinear mapping ability. In this dissertation, we present an extended SDM which can overcome the original SDM poor nonlinear mapping ability by modifying original presetting mode of address matrix and learning algorithm on the basic of former studies.In the ExSDM, binary coding for input data is replaced by direct real-valued inputs, accordingly, avoiding the coding process for data like in SDM. The outer learning rule is replaced by Least Means Squares Error(LMSE), therefore the model has not only the ability of pattern recognition but also the ability of function approximation. The prestting mode of address matrix is automatically decided by sample distribution with the help of (grey) self-organizing feature map ((G)SOFM) so as to reflect the actual sample distribution. The transfer function of the hidden layer is Gaussian function. The computer simulations for non-linear function approximation, time-series prediction and handwritten numeral recognition show that the modified model is effective and feasible.