Evaluation of Latent Space Disentanglement in the Presence of Interdependent Attributes
This addresses the issue of misleading evaluation metrics for researchers and practitioners in music generation, but it is incremental as it builds on existing disentanglement techniques.
The paper tackled the problem of evaluating latent space disentanglement in deep generative models for controllable music generation, where existing metrics like mutual information gap (MIG) are inadequate due to interdependent semantic attributes in real-world music datasets, and they proposed a dependency-aware information metric as a drop-in replacement for MIG.
Controllable music generation with deep generative models has become increasingly reliant on disentanglement learning techniques. However, current disentanglement metrics, such as mutual information gap (MIG), are often inadequate and misleading when used for evaluating latent representations in the presence of interdependent semantic attributes often encountered in real-world music datasets. In this work, we propose a dependency-aware information metric as a drop-in replacement for MIG that accounts for the inherent relationship between semantic attributes.