Enhancing few-shot lifelong learning through fusion of cross-domain knowledge

Humans can continually solve new problems with a few examples and enhance their learned knowledge by incorporating new ones. Few-shot lifelong learning (FSLL) has been presented to mimic human learning ability. However, they overlook the significance of cross-domain knowledge and little effort has been made to investigate it. In this paper, we explore the effects of cross-domain knowledge in FSLL and propose a new framework to enhance the model’s ability by fusing cross-domain knowledge into the learning process. Moreover, we investigate the impact of both debiased and non-debiased models in the FSLL context for the first time. Compared with previous works, our setting presents a unique challenge: the model should continually learn new knowledge from cross-domain few-shot data and update its existing knowledge by fusing new knowledge throughout its lifelong learning process. To address this challenge, the proposed framework focuses on learning and updating while migrating the well-known issues of forgetting and overfitting. The framework comprises three key components designed for learning cross-domain knowledge: the Debiased Base Learning strategy, Knowledge Acquisition, and Knowledge Update. The superiority of the framework is validated on mini-ImageNet, CIFAR-100, OfficeHome, and Meta-Dataset. Experiments show that the proposed framework exhibits the capability to perform in cross-domain situations and also achieves state-of-the-art performance in the non-cross-domain situation.