| 摘要 |
An architecture and data processing method for a neural network that can approximate any mapping function between the input and output vectors without the use of hidden layers. The data processing is done at the sibling nodes (second row). It is based on the orthogonal expansion of the functions that map the input vector to the output vector. Because the nodes of the second row are simply data processing stations, they remain passive during training. As a result the system is basically a single-layer linear network with a filter at its entrance. Because of this it is free from the problems of local minima. The invention also includes a method that reduces the sum of the square of errors over all the output nodes to zero (0.000000) in fewer than ten cycles. This is done by initialization of the synaptic links with the coefficients of the orthogonal expansion. This feature makes it possible to design a computer chip which can perform the training process in real time. Similarly, the ability to train in real time allows the system to retrain itself and improve its performance while executing its normal testing functions. Because the second synaptic link values represent the frequency spectrum of the signal appearing on a given output node, by training the ONN with all N sibling nodes and using only some of them in testing, we can create a low pass, a high pass or a band pass filter.
|
