MMT: Cross Domain Few-Shot Learning via Meta-Memory Transfer

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) typically assumes data from the same distribution for training and testing.
• Domain shift significantly degrades FSL performance when training and testing data originate from different distributions.
• Addressing the cross-domain few-shot learning (CD-FSL) challenge is crucial for real-world applicability.

Purpose of the Study:

• To develop a novel method for effective cross-domain few-shot learning.
• To mitigate the performance drop caused by domain shift in few-shot recognition tasks.
• To propose a Meta-Memory scheme that bridges the gap between source and target domains.

Main Methods:

• Proposed a Meta-Memory scheme incorporating style-memory and content-memory components.
• Style-memory captures intra-domain style information for richer feature distributions.
• Content-memory stores semantic information by exploring diverse category knowledge.
• Employed a contrastive learning strategy to enhance cross-domain adaptation.

Main Results:

• Achieved state-of-the-art performance on cross-domain few-shot semantic segmentation benchmarks (COCO-2ⁱ, PASCAL-5ⁱ, FSS-1000, SUIM).
• Demonstrated positive impact on few-shot classification tasks using the Meta-Dataset.
• Successfully alleviated the cross-domain problem in few-shot learning scenarios.

Conclusions:

• The Meta-Memory scheme effectively bridges domain gaps in few-shot learning.
• The proposed approach offers a robust solution for cross-domain few-shot recognition tasks.
• This work advances the capabilities of few-shot learning in diverse and challenging environments.