Session-Guided Attention in Continuous Learning With Few Samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot class-incremental learning (FSCIL) faces challenges with catastrophic forgetting and limited adaptation to new data due to small sample sizes.
• Current methods often use fixed backbones to preserve old knowledge, but this hinders learning optimal representations for new classes.

Purpose of the Study:

• To propose a novel SEssion-Guided Attention (SEGA) framework to address the limitations of existing FSCIL approaches.
• To improve both knowledge retention and adaptation capabilities in incremental learning scenarios.

Main Methods:

• SEGA leverages class relationships within sessions to assess test sample similarity to class prototypes, enabling accurate session identification.
• An attention module is introduced for each session to refine features from a fixed backbone, enhancing sample clustering within the identified session.

Main Results:

• Experimental results on three FSCIL datasets demonstrate SEGA's superior adaptability.
• The framework effectively avoids forgetting old knowledge while achieving novel data adaptation.

Conclusions:

• The SEGA framework offers a significant advancement in few-shot class-incremental learning.
• SEGA provides a robust solution for scenarios requiring continuous learning with limited data per class.