Balancing Feature Alignment and Uniformity for Few-Shot Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Few-Shot Learning (FSL) aims to recognize new classes with limited data.
• Current FSL methods risk 'supervision collapse' due to base class bias.
• This bias hinders the learning of generalized models for novel classes.

Purpose of the Study:

• To propose a solution addressing the 'supervision collapse' in Few-Shot Learning.
• To develop a method that preserves intrinsic data structure for better generalization.
• To enable the learning of a generalized model applicable to novel classes.

Main Methods:

• The approach maximizes two types of mutual information (MI) based on the InfoMax principle.
• MI is maximized between samples and their feature representations, and between representations and class labels.
• A unified framework uses two low-bias estimators to perturb feature embeddings, combining knowledge distillation and feature diversity.

Main Results:

• The proposed method achieves a balance between discrimination and generalization in feature representations.
• Experimental results on miniImageNet and CIFAR-FS datasets show comparable performance to state-of-the-art methods.
• Achieved 69.53% accuracy on miniImageNet and 77.06% on CIFAR-FS for the 5-way 1-shot task.

Conclusions:

• The proposed Few-Shot Learning approach effectively mitigates 'supervision collapse'.
• The method enhances model generalization capabilities for recognizing novel classes.
• The approach demonstrates strong performance on established FSL benchmarks.