MAGA: Modeling a Group Action

Geonho Hwang, Jaewoong Choi, Hyunsoo Cho, Myungjoo Kang

Keywords: Generative Model, Generalization, Deep Learning, Representation Learning

TL;DR: We make a new generative model that is capable of combinatorial generalization.

Abstract:

Combinatorial generalization, an ability to collect various attributes from diverse data and assemble them to generate novel unexperienced data, is considered an essential traversal point to achieve human-level intelligence.
Previous unsupervised approaches mainly focused on learning the disentangled representation, such as the variational autoencoder. However, recent studies discovered that the disentangled representation is insufficient for combinatorial generalization and is not even correlated. In this regard, we proposed a novel framework of data generation that can robustly generalize under these distribution shift situations. The model, simulating the group action, carries out combinatorial generalization by discovering the fundamental transformation between the data. We conducted experiments on the two settings: Recombination-to-Element, and Recombination-to-Range. The experiments demonstrated that our method has quantitatively and qualitatively superior generalizability and generates better images over traditional models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning