An competitive learning pulsed neural network for temporal signals

In this study, we propose a new competitive learning method for temporal signals using pulsed neuron model. The pulsed neuron models deal with pulse trains as the inputs and outputs, and employ leaky integrators as there internal potentials. Therefore, the models can deal with temporal signals without the windowing process. The proposed method is based on a winner selection method controlling the firing threshold of competitive neurons using a few observer neurons. By employing this method, the winner neuron switches dynamically according to variation of input signals. As a result of the experiment, it become clear that the temporal input signals generated from a real sound could be quantized and the reference vector changes according to variation of input signals.