PatchMix Augmentation to Identify Causal Features in Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to generalize from limited data, crucial for real-world applications.
• Existing FSL methods often overlook distributional shifts caused by sample selection bias.
• This bias can lead to spurious correlations, hindering generalization to novel classes.

Purpose of the Study:

• To address the challenge of sample selection bias in Few-shot Learning.
• To develop a method that identifies and leverages invariant causal features for improved generalization.
• To enhance the discriminative power of learned features for classification tasks.

Main Methods:

• Proposing PatchMix, a novel data augmentation strategy that replaces image patches with data from different classes to break spurious correlations.
• Introducing Correlation-guided Reconstruction (CGR) and a Hardness-Aware module to improve feature discrimination.
• Demonstrating adaptability to unsupervised FSL scenarios.

Main Results:

• PatchMix effectively identifies causal features invariant to distribution shifts.
• The proposed framework achieves state-of-the-art results across multiple benchmarks (miniImageNet, tieredImageNet, CIFAR-FS, etc.) in various FSL settings.
• Quantitative and qualitative analyses confirm the effectiveness of learning causal features.

Conclusions:

• The proposed method successfully mitigates sample selection bias in FSL.
• PatchMix and associated modules enhance model generalization by focusing on causal, invariant features.
• The framework offers a robust solution for both supervised and unsupervised Few-shot Learning tasks.