FS-BAN: Born-Again Networks for Domain Generalization Few-Shot Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot classification (FSC) addresses limited labeled data for novel class recognition.
• Domain generalization FSC (DG-FSC) extends FSC to unseen domains, facing challenges from domain shift.
• Existing models struggle with the domain discrepancy between training (base) and evaluation (novel) classes in DG-FSC.

Purpose of the Study:

• To propose and evaluate novel methods for Domain Generalization Few-Shot Classification (DG-FSC).
• To investigate the effectiveness of Born-Again Network (BAN) episodic training for DG-FSC.
• To introduce Few-Shot BAN (FS-BAN) with specialized multi-task objectives to tackle overfitting and domain discrepancy.

Main Methods:

• Born-Again Network (BAN) episodic training, a knowledge distillation technique, was adapted for DG-FSC.
• Developed Few-Shot BAN (FS-BAN) incorporating novel multi-task learning objectives: Mutual Regularization, Mismatched Teacher, and Meta-Control Temperature.
• Conducted comprehensive quantitative and qualitative analyses across six datasets and three baseline models.

Main Results:

• BAN episodic training shows promise in mitigating domain shift for DG-FSC.
• FS-BAN consistently enhances the generalization performance of baseline models.
• FS-BAN achieves state-of-the-art accuracy on DG-FSC tasks.

Conclusions:

• FS-BAN effectively addresses key DG-FSC challenges like overfitting and domain discrepancy.
• The proposed multi-task objectives are crucial for FS-BAN's superior performance.
• FS-BAN represents a significant advancement in robust few-shot learning across diverse domains.