Deep Mixed Activated Long Short-Term Memory-Restricted Boltzmann Recurrent Neural Network for Piano Melody Generation

Abstract

This paper deals with,a very unique way of implementing the Long Short Term Memory Algorithm (LSTM) in the field of musical melody generation. We have used this LSTM in an advanced way which can be amalgamated with another special deep learning architecture, known as the Restricted Boltzmann Machine (RBM), thus giving rise to a completely new form of LSTM Algorithm for Musical Melody Generation, which we call the Deep Mixed Activated Long Short Term Memory Restricted Boltzmann Recurrent Neural Network Machine (DMA-LSTM-RBRNN). The DMA-LSTM-RBRNN can generate melodies with a corresponding training accuracy of 80.78%, validation accuracy of 81.52% and corresponding training loss of 0.5965, validation loss of 0.5634, and we see that the DMA-LSTM-RBRNN's validation accuracy is leading the training accuracy and the validation loss is lagging the training loss.