Application of Neural Networks to Solve the Task of Generating Music

The article considers the possibility of using neural networks as the basis of the music generator. Under the music generation is the process of creating sequences of music notes, which relates this task to the tasks of simulation sequences, as well as modeling the language. The urgency of the development of this topic is connected with the necessity of writing new music and finding new motives, caused by the production of many films and people's needs. The main requirement for the system was the generation of polyphonic music, a sequence of musical tuples, that is, a multi-dimensional object requiring the use of generative models, such as restricted Boltzman machine, that are able to calculate the probabilistic distribution, by estimating the energy function of configuration state, in relation to recurrent patterns, which in turn enable their production of long-term dependencies in a sequence of multi-dimensional objects. And for the representing of deep relations and dependencies between music notes there were deep belief network used that is able to estimate deep connections and dependencies by hierarchical representing of data. This approach allows to create four topologies which are compound of recurrent part and generative. For this purpose in the article necessary the or ethical information is given. The existing solutions are considered and the iradvantages and disadvantages are determined. The variants of possible topologies, algorithms of the ir training and generation are analyzed and presented. The results of experimental researches on the basis of which the comparisonis made and the advantages and disadvantages of each of the mare determined. For the comparison there were used the negative log-likelihood function, cross-entropy cost, accuracy and training time. The results show that the best characteristics has topology that consists of long short-term memory and deep belief network, because of it difficult structure and properties. But at the same time, training time of this topology was extremely high on small dataset, that makes not efficient to use it as a basis of music generator. Topology which consists of recurrent neural network and restricted Boltzman machine had the worst indicators, but the small training time. Thereby topology which consists of long short-term memory and restricted Boltzman machine had become the best candidate. Good characteristics of this structure are confirmed by the main property of long short-term memory that is solution of vanishing gradient problem and by the possibility of representing complicated distribution in restricted Boltzman machine. Training were performed on “Nottingham folk music database” dataset through 200 epochs. All training and generating algorithms were implemented in Python with Theano. Research described in this article allows to simplify the process of development of polyphonic music generator. Ref. 11, fig. 8, tabl. 1.