Sculpting the Latent Space: Disentanglement with WAE-MMD and Programmable Priors

Name: Sculpting the Latent Space: Disentanglement with WAE-MMD and Programmable Priors
Uploaded: 2026-10-23T12:43:05.133212+00:00
Duration: 812 s
Description: This work introduces a Programmable Prior Framework using Maximum Mean Discrepancy (MMD) to sculpt latent spaces. It allows users to explicitly control the latent distribution, achieving state-of-the-art mutual independence and disentanglement without losing reconstruction quality. The method uses a novel metric, Latent Predictability Score (LPS), to quantify independence reliably.

This work introduces a Programmable Prior Framework using Maximum Mean Discrepancy (MMD) to sculpt latent spaces. It allows users to explicitly control the latent distribution, achieving state-of-the-art mutual independence and disentanglement without losing reconstruction quality. The method uses a novel metric, Latent Predictability Score (LPS), to quantify independence reliably.

Open the full episode