On Generalizing Beyond Domains in Cross Domain Continual Learning (arXiv)

/in /by

Publication Date: 11/19/2021

Event: arXiv

Reference: https://arxiv.org/abs/2203.03970

Authors: Christian Simon, NEC Laboratories America, Inc., The Australian National University, ACT, Australia, Monash University, Masoud Faraki, NEC Laboratories America, Inc., Yi Hsuan Tsai, NEC Laboratories America, Inc., Phiar Technologies, Xiang Yu, NEC Laboratories America, Inc., Samuel Schulter, NEC Laboratories America, Inc., Yumin Suh, NEC Laboratories America, Inc., Mehrtash Harandi, Monash University, VIC, Australia, Data61, Manmohan Chandraker, NEC Laboratories America, Inc., University of California, San Diego

Abstract: Humans have the ability to accumulate knowledge of new tasks in varying conditions, but deep neural networks often suffer from catastrophic forgetting of previously learned knowledge after learning a new task. Many recent methods focus on preventing catastrophic forgetting under the assumption of train and test data following similar distributions. In this work, we consider a more realistic scenario of continual learning under domain shifts where the model must generalize its inference to an unseen domain. To this end, we encourage learning semantically meaningful features by equipping the classifier with class similarity metrics as learning parameters which are obtained through Mahalanobis similarity computations. Learning of the backbone representation along with these extra parameters is done seamlessly in an end to end manner. In addition, we propose an approach based on the exponential moving average of the parameters for better knowledge distillation. We demonstrate that, to a great extent, existing continual learning algorithms fail to handle the forgetting issue under multiple distributions, while our proposed approach learns new tasks under domain shift with accuracy boosts up to 10% on challenging datasets such as DomainNet and OfficeHome.

Publication Link: https://arxiv.org/pdf/2203.03970.pdf

You might also like
5 Accepted Papers CVPR2022! Check out our Computervision Papers at #CVPR2022
5 Accepted Papers CVPR2022! Active Adversarial Domain Adaptation
5 Accepted Papers CVPR2022! Adaptation Across Extreme Variations using Unlabeled Bridges
5 Accepted Papers CVPR2022! Adversarial Learning of Privacy-Preserving and Task-Oriented Representations
5 Accepted Papers CVPR2022! Channel Recurrent Attention Networks for Video Pedestrian Retrieval
5 Accepted Papers CVPR2022! Controllable Dynamic Multi-Task Architectures
5 Accepted Papers CVPR2022! Coordinated Joint Multimodal Embeddings for Generalized Audio-Visual Zero-shot Classification and Retrieval of Videos
5 Accepted Papers CVPR2022! Cross-Domain Similarity Learning for Face Recognition in Unseen Domains