Learning Group Structure and Disentangled Representations of Dynamical Environments

Abstract

Learning disentangled representations is a key step towards effectively discovering and modelling the underlying structure of environments. In the natural sciences, physics has found great success by describing the universe in terms of symmetry preserving transformations. Inspired by this formalism, we propose a framework, built upon the theory of group representation, for learning representations of a dynamical environment structured around the transformations that generate its evolution. Experimentally, we learn the structure of explicitly symmetric environments without supervision from observational data generated by sequential interactions. We further introduce an intuitive disentanglement regularisation to ensure the interpretability of the learnt representations. We show that our method enables accurate long-horizon predictions, and demonstrate a correlation between the quality of predictions and disentanglement in the latent space.

arXiv commentAccepted to NeurIPS 2020

References in corpus (19)

• Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations
• Disentangling by Factorising
• Hamiltonian Neural Networks
• Towards a Definition of Disentangled Representations
• Understanding disentangling in $\beta$-VAE
• DARLA: Improving Zero-Shot Transfer in Reinforcement Learning
• A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning
• Are Disentangled Representations Helpful for Abstract Visual Reasoning?
• Learning Independent Causal Mechanisms
• Independently Controllable Factors
• Learning Compositional Koopman Operators for Model-Based Control
• Learning the Irreducible Representations of Commutative Lie Groups
• A Survey of Inductive Biases for Factorial Representation-Learning
• Hamiltonian Generative Networks
• Decoupling feature extraction from policy learning: assessing benefits of state representation learning in goal based robotics
• Learning Symmetries of Classical Integrable Systems
• Disentangled State Space Representations
• Flatland: a Lightweight First-Person 2-D Environment for Reinforcement Learning
• Understanding image motion with group representations

Cited by in corpus (8)

• Commutative Lie Group VAE for Disentanglement Learning
• Addressing the Topological Defects of Disentanglement via Distributed Operators
• Disentangling by Subspace Diffusion
• Group Equivariant Conditional Neural Processes
• Linear Disentangled Representations and Unsupervised Action Estimation
• A Metric for Linear Symmetry-Based Disentanglement
• Quantifying and Learning Linear Symmetry-Based Disentanglement
• GEM: Group Enhanced Model for Learning Dynamical Control Systems