CSM-Net: Relation embedding for few shot learning optimized by cross memory attention.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-shot learning trains models with minimal labeled data, facing challenges in representation learning and sample relationship modeling.
• Data perturbations significantly impact model performance in few-shot scenarios.

Purpose of the Study:

• To propose Cross-Memory Attention (CMA) for improved few-shot learning.
• To address challenges in learning class-specific representations and modeling sample relationships under limited supervision.
• To mitigate the impact of data perturbations in few-shot learning settings.

Main Methods:

• Cross-Memory Attention (CMA) integrates support and query set memory features to model long-range dependencies.
• A Domain Adaptation Module is introduced to reduce data perturbation impact by training with perturbed data.
• A Multi-sample Adaptive Fusion Module enables CMA application in 5-shot scenarios by extracting common features from multiple samples.

Main Results:

• CMA effectively models relationships between support and query sets with fewer parameters than standard Transformers.
• The Domain Adaptation Module constructs a learnable classification space, improving upon fixed metric learning.
• Extensive experiments on four public datasets demonstrate the model's effectiveness.

Conclusions:

• CMA offers a robust solution for few-shot learning, enhancing generalization and mitigating data perturbation effects.
• The proposed modules provide versatility and adaptability for various few-shot learning tasks and scenarios.