A Cross-Domain Few-Shot Learning Method Based on Domain Knowledge Mapping
Jiajun Chen, Hongpeng Yin, Yifu Yang
TL;DR
This work tackles cross-domain few-shot learning under non-i.i.d. conditions with disjoint label spaces between source and target. It introduces a two-pronged approach: (i) maximum mutual information mixed-supervision pretraining that jointly optimizes $I_{\phi}(\phi(x); Y)$ and $I_{\phi}(\phi(x); X)$ with a dynamic weight $\alpha(t)$ to prevent mode collapse and retain label-agnostic information, and (ii) a domain-knowledge mapping module driven by a domain classifier and gradient-reversal to adapt feature mappings to transfer difficulty, including pseudo-unseen domain generation via Gaussian-noise mixing. The framework is applied consistently across pretraining, meta-training, and meta-testing, enabling rapid adaptation to diverse target domains. Experiments on six diverse datasets show the method outperforms strong meta-learning baselines and existing cross-domain approaches, particularly on fine-grained tasks, validating the practical impact of adaptive domain knowledge transfer for few-shot learning under domain shifts. The work provides a principled way to balance task-specific discriminability with cross-domain invariance, improving generalization when label spaces do not overlap between domains.
Abstract
In task-based few-shot learning paradigms, it is commonly assumed that different tasks are independently and identically distributed (i.i.d.). However, in real-world scenarios, the distribution encountered in few-shot learning can significantly differ from the distribution of existing data. Thus, how to effectively leverage existing data knowledge to enable models to quickly adapt to class variations under non-i.i.d. assumptions has emerged as a key research challenge. To address this challenge, this paper proposes a new cross-domain few-shot learning approach based on domain knowledge mapping, applied consistently throughout the pre-training, training, and testing phases. In the pre-training phase, our method integrates self-supervised and supervised losses by maximizing mutual information, thereby mitigating mode collapse. During the training phase, the domain knowledge mapping layer collaborates with a domain classifier to learn both domain mapping capabilities and the ability to assess domain adaptation difficulty. Finally, this approach is applied during the testing phase, rapidly adapting to domain variations through meta-training tasks on support sets, consequently enhancing the model's capability to transfer domain knowledge effectively. Experimental validation conducted across six datasets from diverse domains demonstrates the effectiveness of the proposed method.