Variational Learning of Disentangled Representations

ICLR 2026 Conference Submission22598 Authors

Keywords: Disentangled representation learning, domain invariance, variational inference

TL;DR: This work proposes a variational framework that disentangles shared and condition-specific factors using a dual-latent architecture and reconstructions, allowing for invariant representation learning.

Abstract:

Disentangled representations allow models to separate factors shared across conditions from those that are condition-specific. This separation is crucial in domains such as biomedicine, where generalization to new treatments, patients, or species requires isolating stable biological signals from context-dependent effects. While several VAE-based extensions aim to achieve this, they often exhibit leakage between latent variables, limiting generalization.
We introduce DisCoVR, a variational framework that explicitly separates invariant and condition-specific factors through: (i) a dual-latent architecture, (ii) parallel reconstructions to keep both representations informative, and (iii) a max–min objective that enforces separation without handcrafted priors.
We show this objective maximizes data likelihood, promotes disentanglement, and admits a unique equilibrium.
Empirically, DisCoVR achieves stronger disentanglement on synthetic data, natural images, and single-cell RNA-seq datasets, establishing it as a principled approach for multi-condition representation learning.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 22598