A single-layer perceptron for Fourier transform

Summary form only given, as follows. A single-layer perceptron has been trained to perform the fast Fourier transform. Excellent performance has been found through computer simulations. The peak amplitude of the filter sidelobes can be reduced significantly relative to the peak amplitude of the mainlobe, but the width of the mainlobe does not increase at all. Computer simulations show that the peak amplitude of the first sidelobe is zero for several cases. The single-layer perceptron finishes computation in two stages regardless of the number of samples in the time domain; this implies that the computation speed is fast. From simulation results it has been found that the number of neurons in the hidden layer determines the accuracy of output. More neurons in the hidden layers do not imply better performance. The number of neurons in the hidden layer can be as small as log n/log 2, where n is the number of input signal samples. This implies that the hardware implementation for real-time application is feasible. Results from a preliminary study show that the single-layer perceptron can suppress spectral leakage better than the blackman window, without widening the mainlobe, even with several faulty weights. The backpropagation training algorithm has been used to train the single-layer perceptron with specially designed data