Kryvokhata A. Neural networks mathematical models for the sound signals recognition problems

The thesis is devoted to solving problems of sound classification using convolutional neural networks and autoencoders with optimization of their structure by genetic algorithms. Mathematical models of the hybrid neural networks based on convolutional architecture, autoencoder and Snapshot ensemble method are suggested. The influence of the classifier hyperparameters on the model accuracy is investigated. A convergence theorem is proved for a hybrid neural network with autoencoder and convolution layers. An instrumental system for classifying audio data is implemented. A genetic algorithm is used to improve the ensembles structure and to automatically tune the hyperparameters.