Fully Self-Supervised Out-of-Domain Few-Shot Learning with Masked Autoencoders.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to classify new categories with minimal labeled data.
• Existing FSL methods struggle with out-of-domain generalization and often rely on supervised fine-tuning.
• This reliance limits adaptability and increases data requirements.

Purpose of the Study:

• To propose a novel, fully self-supervised few-shot learning (FSS) technique.
• To enhance generalization capabilities of FSL models, particularly in out-of-domain scenarios.
• To reduce dependency on supervised fine-tuning for FSL tasks.

Main Methods:

• Developed a fully self-supervised few-shot learning technique (FSS).
• Utilized a vision transformer architecture combined with a masked autoencoder.
• Employed episode-wise, fully self-supervised fine-tuning for out-of-domain generalization.

Main Results:

• Achieved accuracy gains of 1.05% (ISIC), 0.12% (EuroSat), and 1.28% (BCCD) on out-of-domain datasets.
• Demonstrated effective generalization without any supervised training.
• Validated the FSS technique across three diverse datasets.

Conclusions:

• The proposed FSS technique successfully generalizes to unseen, out-of-domain classes.
• Fully self-supervised fine-tuning offers a viable alternative to supervised methods in FSL.
• FSS presents a promising direction for robust and adaptable few-shot learning systems.