A Few-Shot Class Incremental Learning Method Using Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot class incremental learning (FSCIL) faces challenges in balancing stability and plasticity.
• Existing FSCIL methods struggle with dynamic learning scenarios and catastrophic forgetting.

Purpose of the Study:

• To develop a novel FSCIL framework that enhances stability and plasticity.
• To improve cross-modal alignment and mitigate catastrophic forgetting in incremental learning.

Main Methods:

• Leveraging graph neural networks (GNNs) to model interdependencies between categories.
• Utilizing a Graph Isomorphism Network (GIN), Hamiltonian Graph Network with Energy Conservation (HGN-EC), and Adversarially Constrained Graph Autoencoder (ACGA).
• Integrating a parameter-efficient CLIP backbone with contrastive learning and energy-based regularization.

Main Results:

• The proposed framework demonstrates improved incremental accuracy and stability on benchmark datasets.
• Validation against state-of-the-art baselines confirms the framework's effectiveness.
• The method successfully models semantic correlations between textual and visual modalities.

Conclusions:

• The novel FSCIL framework unifies graph-based relational reasoning with physics-inspired optimization.
• This approach offers a scalable and interpretable solution for dynamic learning scenarios.
• The work advances the field of FSCIL by addressing key limitations in existing methods.