Domain Generalization In Robust Invariant Representation
Gauri Gupta, Ritvik Kapila, Keshav Gupta, Ramesh Raskar
TL;DR
The paper addresses domain generalization of invariances by modeling identity-preserving transformations as a group action of $G$ on $X$ with orbits $O_x$ and invariant latent $z \in Z = X/G$. It introduces RotInvVAE, which yields an invariant code $z$ and an estimated group action, enabling reconstruction via the decoder; pair verification relies on cosine similarity between $z$ codes. Across unseen domains such as MNIST, FashionMNIST, and LFW, RotInvVAE maintains rotation-invariant latent representations and demonstrates strong generalization, including cross-domain transfers and reduced-label scenarios. The work argues for resource-efficient, domain-general invariance in recognition tasks, offering a first demonstration that invariance can robustly transfer to new domains without retraining.
Abstract
Unsupervised approaches for learning representations invariant to common transformations are used quite often for object recognition. Learning invariances makes models more robust and practical to use in real-world scenarios. Since data transformations that do not change the intrinsic properties of the object cause the majority of the complexity in recognition tasks, models that are invariant to these transformations help reduce the amount of training data required. This further increases the model's efficiency and simplifies training. In this paper, we investigate the generalization of invariant representations on out-of-distribution data and try to answer the question: Do model representations invariant to some transformations in a particular seen domain also remain invariant in previously unseen domains? Through extensive experiments, we demonstrate that the invariant model learns unstructured latent representations that are robust to distribution shifts, thus making invariance a desirable property for training in resource-constrained settings.