CORE: CORrelation-Guided Feature Enhancement for Few-Shot Image Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot classification (FSC) struggles with limited data for novel classes, leading to biased feature distribution estimation and inaccurate decision boundaries.
• Outliers in support data further exacerbate these issues in traditional FSC methods.

Purpose of the Study:

• To propose a novel feature enhancement method, CORrelation-guided feature Enrichment (CORE), to improve few-shot classification performance.
• To address the challenge of inadequate representation of intraclass variations in novel classes during few-shot learning.

Main Methods:

• Developed CORE, a feature enhancement method utilizing an autoencoder (AE) architecture with classification information integrated into its latent space.
• Trained CORE on base classes to leverage its generative ability for novel classes, generating discriminative features while reducing irrelevant content.
• Employed weak supervision from base classes to guide feature generation for novel classes.

Main Results:

• CORE effectively generates improved features for novel classes, reducing estimation bias of the class distribution.
• The method demonstrated reduced sensitivity to the selection of support data in few-shot learning.
• Experiments showed consistent outperformance over existing methods across various benchmarks using different backbones and classifiers.

Conclusions:

• CORE offers a generic and flexible approach to enhance features in few-shot learning, easily integrable with existing methods.
• The proposed method significantly improves the accuracy and robustness of few-shot classification, particularly when dealing with limited and potentially outlier-prone data.
• CORE advances the state-of-the-art in few-shot learning by providing a more reliable feature representation for novel classes.