Music melody generation and LIF supervised training based on spiking neural network

Abstract

With the help of neuroscience research, the neural network model based on the pulse generation time can be accurately established, which is called Spiking Neural Network (SNN). This new type of neural network uses spike coding, which allows for more information and greater computing power to be obtained by obtaining the precise timing of the spikes. The application of SNN in music melody generation has a good application prospect. In this study, the Cell of Long Short-Term Memory (LSTM) in the generative model was replaced by Leaky Integrate-and-Fire (LIF), and two different variants of LIF were proposed based on Cell. Although the LIF variant achieved better results than the original LIF, it was not better than Cell in comprehensive comparison, and replacing Cell with LIF in the generative model could not improve the quality of the generated melody. The simple strategy used in the experiments to solve the supervised training problem of LIF was the main reason for such results. For SNN to be widely used, it is urgent to solve its supervised training problem.