A prompt regularization approach to enhance few-shot class-incremental learning with Two-Stage Classifier.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-Shot Class-Incremental Learning (FSCIL) addresses training models with limited data and continuous learning without catastrophic forgetting.
• Pre-trained models offer robust feature representations and transferability, crucial for both few-shot and incremental learning.
• Prompt Learning enhances pre-trained model performance on downstream tasks, especially in large-scale vision and language models.

Purpose of the Study:

• To propose a novel Prompt Regularization (PrRe) approach for Few-Shot Class-Incremental Learning (FSCIL).
• To enhance pre-trained Vision Transformers (ViTs) by fusing Task and Global Prompts.
• To improve model efficiency and prevent knowledge forgetting in incremental learning scenarios.

Main Methods:

• Developed a Prompt Regularization (PrRe) method embedding Task and Global Prompts within a pre-trained Vision Transformer (ViT).
• Introduced a Two-Stage Classifier (TSC) using K-Nearest Neighbors for base sessions and a Prototype Classifier for incremental sessions.
• Integrated a global self-attention module within the classification phase.

Main Results:

• Demonstrated the effectiveness of the proposed PrRe method through experiments on multiple benchmark datasets.
• Showcased the superiority of the PrRe approach in Few-Shot Class-Incremental Learning tasks.
• Validated the ability of the method to efficiently update models without forgetting previously learned information.

Conclusions:

• The proposed Prompt Regularization (PrRe) approach effectively enhances pre-trained Vision Transformers for FSCIL.
• The Two-Stage Classifier (TSC) integrated with global self-attention proves effective for incremental learning sessions.
• The method offers a promising solution for efficient and stable incremental learning with limited data.