Few-Shot Continual Learning via Flat-to-Wide Approaches.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Traditional continual learning (CL) methods require extensive data, limiting their application in real-world scenarios with sample scarcity.
• Overfitting is a significant challenge in CL when dealing with limited datasets.
• Catastrophic forgetting (CF) remains a key problem in sequential learning tasks.

Purpose of the Study:

• To propose a few-shot continual learning (CL) approach, named FLOWER (flat-to-wide approach).
• To address the limitations of existing CL methods, particularly concerning data scarcity and catastrophic forgetting (CF).
• To enhance model performance in scenarios with limited training samples.

Main Methods:

• Introduced the flat-to-wide learning process to identify flat-wide minima, mitigating catastrophic forgetting (CF).
• Employed a data augmentation strategy utilizing a ball-generator concept to constrain the sampling space within the smallest enclosing ball, tackling data scarcity.
• Developed and evaluated the FLOWER approach on benchmark tasks.

Main Results:

• FLOWER demonstrated significantly improved performance compared to existing methods, especially on small base tasks.
• The proposed approach effectively overcomes the challenge of data scarcity in continual learning.
• The flat-to-wide learning process successfully addressed the catastrophic forgetting (CF) problem.

Conclusions:

• FLOWER presents a viable solution for few-shot continual learning (CL) challenges.
• The method offers a practical alternative for real-world applications with limited data.
• The findings highlight the effectiveness of the flat-to-wide minima and ball-generator augmentation for CL.