Disentanglement by means of action-induced representations

2026-02-06

AI Summary

Learning interpretable and disentangled representations—where each latent dimension corresponds to an independent generative factor—remains a central challenge in unsupervised representation learning. This work proposes the Action-Induced Representation (AIR) framework, which explicitly links interventions (actions) to the affected degrees of freedom by modeling how physical systems respond under different actions. Built upon a variational autoencoder architecture (VAIR), this approach provides the first theoretical guarantee that generative factors can be provably disentangled under action-based interventions. Experiments demonstrate that VAIR successfully learns disentangled representations even in scenarios where standard VAEs fail, and accurately captures the dependency of each generative factor on specific actions.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Knowledge Representation and Reasoning: Action, Change, and Causality
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Learning interpretable representations with variational autoencoders (VAEs) is a major goal of representation learning. The main challenge lies in obtaining disentangled representations, where each latent dimension corresponds to a distinct generative factor. This difficulty is fundamentally tied to the inability to perform nonlinear independent component analysis. Here, we introduce the framework of action-induced representations (AIRs) which models representations of physical systems given experiments (or actions) that can be performed on them. We show that, in this framework, we can provably disentangle degrees of freedom w.r.t. their action dependence. We further introduce a variational AIR architecture (VAIR) that can extract AIRs and therefore achieve provable disentanglement where standard VAEs fail. Beyond state representation, VAIR also captures the action dependence of the underlying generative factors, directly linking experiments to the degrees of freedom they influence.

Problem

• disentanglement
• representation learning
• variational autoencoders
• independent component analysis
• interpretable representations

Innovation

• disentanglement
• action-induced representations
• variational autoencoders
• representation learning
• generative factors