For more details, we invite you to read the full paper, or come talk to us at ICML 2022! 10/10 🧵

3) LSBD representations also satisfy previous disentanglement notions: Good scores for D_LSBD (lower is better) typically imply good scores on traditional disentanglement metrics (higher is better), even if the reverse is not true. 9/

2) LSBD-VAE and other LSBD methods *can* learn LSBD representations with limited supervision on transformations: Methods that specifically target LSBD indeed score better on D_LSBD. 8/

1) Traditional disentanglement methods don't learn LSBD representations: Methods without a notion of the underlying symmetry group structure don't score well on D_LSBD, even if they do score well on traditional disentanglement metrics. 7/

We perform experiments on a number of datasets with underlying symmetries, using our own LSBD-VAE and other LSBD-oriented methods, as well as various traditional disentanglement methods. From this we draw 3 main conclusions: 6/

We show an intuitive justification of our D_LSBD metric, as well as its theoretical derivation. We also provide a practical implementation to compute D_LSBD for SO(2) symmetry groups of 2D rotations. 5/

In our paper we present the following contributions: i) D_LSBD, a well-formalized general metric to quantify LSBD in learned representations. ii) LSBD-VAE, a weakly-supervised method to learn LSBD representations under certain assumptions. 4/

However, there is currently no metric to quantify LSBD. Such a metric is crucial to evaluate methods aimed at learning LSBD, and to compare to previous notions of disentanglement. 3/

Disentanglement is an important goal in representation learning, but there is no consensus on a formal definition. Higgins et al. (2018) propose such a definition based on the idea that real-world symmetries provide a structure to disentangle, which we refer to as LSBD. 2/

Our paper "Quantifying and Learning Linear Symmetry-Based Disentanglement" got accepted at #ICML2022 @icmlconf!🥳🎉 Congrats to co-authors @tezcatlipocta (joint first), @vlamen, Mike Holenderski & Jim Portegies. Paper: arxiv.org/abs/2011.06070 Code: github.com/luis-armando-pere… 1/🧵