PKI: Prior knowledge-infused neural network for few-shot class-incremental learning

2025-07-10

Neural Networks

AI Summary

This work addresses catastrophic forgetting and overfitting to novel classes in few-shot class-incremental learning by proposing a Prior Knowledge Integrated (PKI) neural network. The core innovation lies in a cascaded projector mechanism: at each incremental stage, a new projector is introduced and cascaded with historical projectors, while the backbone network remains frozen and only the new projector and classifier are fine-tuned, thereby dynamically integrating multi-stage prior knowledge. Additionally, lightweight variants PKIV-1 and PKIV-2 are designed to balance performance and efficiency. Extensive experiments on three mainstream benchmarks demonstrate that the proposed method significantly outperforms existing approaches, effectively mitigating forgetting and enhancing the model’s ability to learn new classes.

Technology Category

• Machine Learning: Multi-instance/Multi-view Learning
• Search and Optimization: Learning to Search
• Computer Vision: Learning & Optimization for CV

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Problem

• few-shot class-incremental learning
• catastrophic forgetting
• overfitting
• prior knowledge

Innovation

• few-shot class-incremental learning
• prior knowledge infusion
• projector ensemble
• catastrophic forgetting mitigation
• parameter-efficient adaptation