抄録
The multilayer perceptron called Sandglass type Neural Network (SNN) has the same number of units in input layer and output layer and has less units in hidden layer than units both in input layer and output layer. In this paper we clarified the properties of a noise reduction filter using the SNN. The properties were derived basically by use of the result that the output signal of the SNN could be given by Karhunen-Loeve expansion of an input data matrix. Here, we evaluated the improvement value of signal to noise ratio for the optimum number of hidden units. The noise reduction filter was assured to be effective and stable by the computer experiments using sinusoidal signal corrupted by white noise.
