Few-Shot Learning With a Strong Teacher.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Few-shot learning (FSL) aims to build classifiers from limited labeled data.
• Meta-learning approaches train a model on multiple tasks to adapt to new tasks with few examples.
• Existing meta-learning for FSL faces challenges with insufficient supervision from sampled query examples and diminishing effectiveness with more shots.

Purpose of the Study:

• To address weaknesses in current meta-learning for FSL.
• To propose a novel meta-training objective that encourages the generation of strong classifiers.
• To improve the performance and applicability of few-shot learning methods.

Main Methods:

• Introduced a novel meta-training objective using 'strong classifiers' trained on ample data.
• Used strong classifiers to supervise the few-shot learner, guiding it to generate better classifiers.
• Integrated the proposed objective as a plug-and-play term into existing meta-learning FSL methods.

Main Results:

• Demonstrated notable improvements across various tasks on benchmark datasets like miniImageNet and tieredImageNet.
• Showcased consistent outperformance of meta-learning FSL methods over non-meta-learning methods, even in many-shot scenarios.
• Validated the approach's effectiveness when combined with representative meta-learning techniques.

Conclusions:

• The proposed 'Learning with A Strong Teacher' approach significantly enhances meta-learning based few-shot learning.
• The method improves classifier performance and broadens the applicability of FSL, particularly in settings with more available data per class.
• This work provides a robust and easily integrated solution for advancing few-shot classification capabilities.