Exploring Convolutional Neural Networks for Music Information Retrieval: Neural Architecture Search, Data Augmentation, and Explainability

This thesis explores three key aspects of applying Convolutional Neural Networks (CNNs) to Music Information Retrieval (MIR): hyperparameter optimization, data augmentation, and explainability. First, I developed a custom Neural Architecture Search method based on genetic algorithms to optimize CNNs for classifying guitar effect chains, achieving a better balance between accuracy and model compactness than Random Search. Second, I investigated data augmentation for Music Emotion Recognition (MER) on guitar recordings, systematically testing 11 techniques and showing that pitch shifting, time stretching, and time shifting were the most effective without significantly affecting perceived emotion. Finally, I studied explainability in MER by adapting Grad-CAM, SHAP, and LIME to musical spectrograms, and I developed an application that provides multi-level explanations of emotion predictions in guitar improvisations.