Evolutionary Deep Learning for Sequential Data Processing in Music Education

Abstract

This paper presents an in-depth study and analysis of music education using evolutionary deep learning algorithms for sequential data processing and designs corresponding models for application in the actual music teaching process. By studying structured composition, a structured model based on motive-phrase-phrase and phrase is proposed for other automatic composition methods that are learned with the whole piece of music, resulting in the shortcoming of insufficient musical structure, starting from the composition structure of motive-phrase-phrase, and using deep learning techniques to learn composition. In the music generation model, a Scratch music generation model capable of generating music in Pianoroll format is constructed by using a generative adversarial network based on sentiment and temporal structure, and a convolutional neural network is used in the generator and discriminator to enhance the training speed. The two algorithmic models are validated by multiple sets of comparison experiments and algorithmic validity experiments to verify the effectiveness and practicality of the models. The method achieves structural feature extraction of music by designing feature extractors on different music granularities. By designing the feature expression function on multi-scale music granularity, the music structure embedded in the music itself is incorporated into the reward function. The results of the subjective and objective evaluation indexes of the experiments show that the method can achieve better music generation results than the manual rule-based and backward and forward relationship-based reward function methods, solving the problem of lack of music theory knowledge to propose rules and compensating the pain of insufficient use of structural information of music created by the backward and forward relationship-based network model. The parameters of the model are updated with the help of the forward-backward propagation method, and the dropout technique is used to improve the overfitting resistance of the model. The test results show that the model has a specific generalization ability, the correct rate can reach 90%, and the model’s recall rate and accuracy rate are high.