Reconstruction for disentanglement, Contrast for invariance

Jiageng Zhu, Hanchen Xie, Wael AbdAlmageed

Keywords: Representation learning, Disentanglement learning, Invariant learning

Abstract: Disentangled and invariant representation are two vital goals for representation learning and many approaches have been proposed to achieve one of them. However, those two goals are actually complementary to each other and we propose a framework to accomplish both of them together. We introduce weakly supervised signals to learn disentangled representation and use contrastive methods to enforce invariant representation. By experimenting on state-of-the-art datasets, the results show that our framework outperforms previous works on both tasks.

One-sentence Summary: Achieve disentanglement representation and invariant representation at the same time.