Mutual Correlation Network for few-shot learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Metric-based Few-Shot Learning (FSL) methods primarily focus on learning effective image embeddings.
• Existing FSL approaches often struggle to capture cross-correlations between image pairs or are limited by the receptive field of Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To address limitations in current FSL methods by proposing a novel network architecture.
• To improve the exploration of cross-correlations and global consensus within image feature maps for enhanced few-shot classification.

Main Methods:

• Introduction of the Mutual Correlation Network (MCNet), incorporating a self-attention mechanism for a global receptive field.
• MCNet features a multi-level embedding module for hierarchical semantic capture and a mutual correlation module for refining correlation maps and generating robust relational embeddings.

Main Results:

• MCNet demonstrates competitive performance across four standard few-shot classification benchmarks: miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS.
• The proposed method effectively explores global consensus in correlation maps, overcoming limitations of CNNs' restricted receptive fields.

Conclusions:

• The Mutual Correlation Network (MCNet) offers a significant advancement in metric-based Few-Shot Learning.
• MCNet's ability to leverage self-attention for global correlation analysis leads to improved performance in few-shot image classification tasks.