An extension of disentanglement metrics and its application to voice - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

An extension of disentanglement metrics and its application to voice

Résumé

In representation learning, the promise of disentanglement methods is to decompose an input signal into a set of independent and interpretable attributes. Some metrics, such as the DCI or MIG scores, have been proposed to evaluate how much this goal is reached. They analyse the relationship between the representation components and the desirable attributes. This paper shows that, even when applied to synthetic datasets generated from a closed list of generative factors, these metrics can be too optimistic. In particular, it reports that a generative factor can be recovered from an altered disentangled representation from which it has been supposedly removed, according to the metrics. Based on this observation, a new criterion called latent decimation is proposed to evaluate disentanglement through the accuracy of factors prediction from subsets of latents. A new metric called MIDCI is defined, and its relevance is demonstrated on voice data.
Fichier non déposé

Dates et versions

hal-04449163 , version 1 (09-02-2024)

Identifiants

Citer

Olivier Zhang, Olivier Le Blouch, Nicolas Gengembre, Damien Lolive. An extension of disentanglement metrics and its application to voice. INTERSPEECH 2023, Aug 2023, Dublin, Ireland. pp.2878-2882, ⟨10.21437/Interspeech.2023-383⟩. ⟨hal-04449163⟩
6 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More