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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot class-incremental learning (FSCIL) faces challenges of catastrophic forgetting and overfitting to new classes.
• Existing methods often freeze network components to preserve prior knowledge, potentially limiting adaptation.
• Balancing knowledge retention and new learning is crucial for effective incremental learning.

Purpose of the Study:

• To propose a novel Prior Knowledge-Infused neural network (PKI) for enhanced FSCIL.
• To effectively integrate and utilize accumulated prior knowledge during incremental learning sessions.
• To mitigate catastrophic forgetting and overfitting while improving new class recognition.

Main Methods:

• The PKI model comprises a backbone, an ensemble of projectors, a classifier, and memory.
• A new projector is added and fine-tuned with the classifier in each incremental session.
• Cascading projectors integrate prior knowledge, enabling flexible learning of new information.

Main Results:

• The proposed PKI approach demonstrates superior performance in recognizing both old and new classes.
• Variants PKIV-1 and PKIV-2 offer a trade-off between resource consumption and performance.
• Extensive experiments on three benchmarks show PKI outperforms state-of-the-art FSCIL methods.

Conclusions:

• The PKI network effectively leverages prior knowledge for robust few-shot class-incremental learning.
• The cascading projector design facilitates flexible integration of new knowledge.
• PKI offers a promising direction for continual learning systems facing limited data.