Episodic Training for Domain Generalization Using Latent Domains

Domain generalization (DG) is to learn knowledge from multiple training domain, and build a domain-agnostic model that could be used to an unseen domain. In this paper, take advantage of aggregating data method from all source and latent domains as a novel, we propose episodic training for domain generalization, aim to improve the performance during the trained model used for prediction in the unseen domain. To address this goal, we first designed an episodic training procedure that train a domain-generalized model without using domain labels. Firstly, we divide samples into latent domains via clustering, and design an episodic training procedure. Then, trains the model via adversarial learning in a way that exposes it into domain shift which decompose the model into feature extractor and classifier components, and train each component on the episodic domain. We utilize domain-invariant feature for clustering. Experiments show that our proposed method not only successfully achieves un-labeled domain generalization but also the training procedure improve the performance compared conventional DG methods.