Exploring Variational Auto-Encoder Architectures, Configurations, and Datasets for Generative Music Explainable AI

Explainability in music AI requires deliberate design choices: The paper demonstrates that making generative music AI understandable isn't automatic—researchers must systematically test different model architectures and configurations to find which ones produce interpretable results that humans can meaningfully control.

Simpler music is easier to explain: MeasureVAE performed best when generating simpler musical genres like pop and rock, suggesting that explainable AI systems may have inherent limitations depending on the complexity of what they're trying to create, which has implications for which creative domains can be made transparent.

Model architecture affects different aspects of interpretability differently: While MeasureVAE excelled at reconstructing music accurately, AdversarialVAE performed better at keeping musical attributes independent from each other—highlighting that there's no single "best" approach and trade-offs exist between different explainability goals.

Specific technical parameters matter for real-world deployment: The recommendation that "32 or 64 latent dimensional space is optimal for 4 regularised dimensions" shows that responsible AI implementation requires detailed technical tuning; one-size-fits-all approaches won't work for making generative models interpretable.

Systematic testing of AI configurations is still lacking: The authors note this is "the first detailed comparison" of these model configurations, indicating that the field of explainable generative AI still needs more rigorous, comprehensive research before practitioners can confidently deploy these systems responsibly.