Random rotational embedding Bayesian optimization for human-in-the-loop personalized music generation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Creativity

Background:

• Generative deep learning models create diverse outputs via latent space sampling.
• Personalizing these models requires efficient optimization of user preferences within the latent space.
• Bayesian optimization is a key technique for human-in-the-loop optimization.

Purpose of the Study:

• Introduce Random Rotational Embedding Bayesian Optimization (ROMBO) for efficient high-dimensional optimization.
• Enable personalized content generation by optimizing user queries in generative models.
• Evaluate ROMBO's effectiveness in a music generation task.

Main Methods:

• Developed ROMBO, embedding low-dimensional Gaussian spaces into high-dimensional ones using random rotations.
• Applied ROMBO to optimize queries for a generative deep learning music model.
• Conducted simulated experiments and user studies (n=16) for evaluation.

Main Results:

• ROMBO demonstrated superior performance over baseline methods in high-dimensional optimization.
• Achieved 16%-31% loss reduction in simulated music generation tasks.
• User studies showed a 40% increase in finding favorite music, 16% faster discovery, and 18% less time on disliked music.

Conclusions:

• ROMBO offers an efficient and effective method for personalizing generative deep learning models.
• The approach significantly enhances user experience and satisfaction in content generation tasks.
• ROMBO shows promise for applications requiring sample-efficient optimization in high-dimensional, rotationally symmetric spaces.