A Chalmers University of Technology Master’s thesis template for LATEX: Investigation of Curriculum Learning in Deep Generative Modelling Using Western Classical Music

Abstract

Current deep learning approaches for symbolic music generation typically train on randomly ordered musical sequences, which can hinder the development of coherent musical structure and effective learning of long-term dependencies. While curriculum learning has demonstrated significant benefits in natural language processing and computer vision by progressively introducing task complexity, its application to symbolic music generation remains largely unexplored. The hierarchical nature of Western classical music, with structures ranging from simple motifs to complex compositions, makes it an ideal candidate for progressive learning strategies.

This thesis investigates the effectiveness of curriculum learning strategies for symbolic music generation using the Music Transformer architecture. A loss-based complexity ranking system is implemented to order musical sequences from simple to complex, combined with progressive data exposure schedules. Our experimental framework compares baseline training against three curriculum learning variants (CL-60%, CL-80%, and CL-60% with learning rate adaptation) using MAESTRO dataset of Western classical piano compositions.

While curriculum models demonstrated faster early convergence, they produced broader loss distributions with higher variance compared to the baseline’s concentrated performance. The CL-60% LR variant emerged as a good performer, achieving superior results in multiple musical features including polyphony, qualified note ratio, and rhythmic structure. Importantly, curriculum learning preserved generalization capability while offering enhanced musical expressiveness.

Although curriculum models did not significantly outperform the baseline in loss metrics, they generated outputs that were more harmonically rich, structurally coherent, and aligned with real music characteristics. These findings demonstrate that curriculum learning offers valuable trade-offs in symbolic music generation, producing more musically compelling outputs when properly designed. This work establishes curriculum learning as a promising training paradigm for music generation and highlights the importance of moving beyond loss-based evaluation toward music-informed assessment metrics. All code and experiments are available at: Curriculum_learning 1