Robust Meta-Representation Learning via Global Label Inference and Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to train models with limited labeled data, a key challenge in meta-learning.
• Feature pre-training enhances FSL generalization but lacks theoretical understanding and requires global labels.
• The scarcity of global labels in real-world scenarios limits the applicability of pre-training strategies.

Purpose of the Study:

• To establish a theoretical connection between feature pre-training and meta-learning for FSL.
• To introduce a novel meta-learning algorithm, Meta Label Learning (MeLa), that overcomes the need for global labels.
• To enhance meta-representation learning through an augmented pre-training procedure.

Main Methods:

• The study theoretically analyzes the link between pre-training and meta-learning, explaining the robustness of meta-representations.
• Meta Label Learning (MeLa) is proposed, a method that infers global labels across tasks to enable pre-training.
• An augmented pre-training strategy is introduced to further refine the learned meta-representation.

Main Results:

• MeLa demonstrates superior performance compared to existing methods across various benchmarks.
• The algorithm is particularly effective in challenging settings with limited training tasks and task-specific labels.
• Theoretical analysis supports the empirical findings on pre-training's contribution to generalization.

Conclusions:

• Feature pre-training is crucial for robust meta-representation in FSL.
• MeLa offers a practical solution for leveraging pre-training in FSL, even when global labels are unavailable.
• The proposed methods advance the field of meta-learning and few-shot learning.