Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Learning Group Structure and Disentangled Representations of Dynamical Environments

About

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.

Robin Quessard, Thomas D. Barrett, William R. Clements• 2020

Related benchmarks

TaskDatasetResultRank
Disentanglement3DShapes
DCI Score0.97
22
DisentanglementFlatLand permutation colors
Beta-VAE1
15
DisentanglementCOIL 2
Beta-VAE100
15
DisentanglementCOIL3
Beta-VAE0.96
15
Disentangled Representation LearningFlatLand rotation colors
Beta-VAE Score0.99
15
DisentanglementMPI3D
Modularity Score47
14
Prediction ErrorCOIL2 ood restricted (right-most rotation) (test)
Seen Prediction Error5.10e-5
6
Prediction ErrorCOIL2 iid restricted (test)
Seen Prediction Error1.70e-4
6
DisentanglementCOIL3 (iid)
Beta-VAE98
5
DisentanglementCOIL2 (ood)
Beta-VAE100
5
Showing 10 of 14 rows

Other info

Follow for update