Memorizing Complementation Network for Few-Shot Class-Incremental Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Few-shot Class-Incremental Learning (FSCIL) faces challenges like catastrophic forgetting and overfitting due to limited data.
• Balancing retention of old knowledge with learning new concepts is difficult in FSCIL.

Purpose of the Study:

• To propose a novel method, the Memorizing Complementation Network (MCNet), to enhance FSCIL performance.
• To mitigate catastrophic forgetting and overfitting in incremental learning scenarios.

Main Methods:

• MCNet ensembles multiple models, leveraging their diverse knowledge memorization for new tasks.
• A Prototype Smoothing Hard-mining Triplet (PSHT) loss is developed to refine model updates with few samples.
• PSHT pushes novel samples away from each other and the old data distribution.

Main Results:

• MCNet demonstrated superior performance on benchmark datasets including CIFAR100, miniImageNet, and CUB200.
• The proposed method effectively addresses the trade-off between old knowledge retention and new concept learning.
• Experimental results validate the efficacy of MCNet and PSHT loss in few-shot incremental learning.

Conclusions:

• MCNet offers a promising approach to ensemble learning for improving FSCIL.
• The PSHT loss effectively aids in learning from scarce data while preserving prior knowledge.
• The study highlights the potential of complementary model knowledge for robust incremental learning.