Evaluation of Latent Space Disentanglement in the Presence of Interdependent Attributes

TL;DR

This paper introduces a dependency-aware information metric to better evaluate latent space disentanglement in deep generative models, especially when attributes are interdependent, improving upon existing metrics like MIG.

Contribution

The paper proposes a new metric that accounts for attribute dependencies, addressing limitations of current disentanglement evaluation methods in complex, real-world datasets.

Findings

The new metric provides more accurate disentanglement assessment in interdependent attribute scenarios.

Existing metrics like MIG can be misleading when attributes are correlated.

The dependency-aware metric enhances evaluation robustness for music generation models.

Abstract

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.