Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• The dynamic nature of the real world necessitates AI models that can adapt to new information without losing previously acquired knowledge.
• Few-shot class-incremental learning (FSCIL) addresses the challenge of recognizing novel classes with limited data while preserving performance on existing classes.
• Existing methods struggle with catastrophic forgetting and limited data availability in incremental learning scenarios.

Purpose of the Study:

• To introduce a novel meta-learning paradigm, Limit (LearnIng Multi-phase Incremental Tasks), for few-shot class-incremental learning.
• To develop a method that efficiently adapts to new classes and mitigates forgetting of old classes.
• To establish a robust framework for recognizing novel classes with minimal data while maintaining performance on established classes.

Main Methods:

• Limit synthesizes fake FSCIL tasks from a base dataset to create a meta-learning environment.
• A transformer-based calibration module aligns old and new class representations, bridging semantic gaps.
• Instance-specific embeddings are adaptively contextualized using a set-to-set function within the calibration module.

Main Results:

• Limit demonstrates state-of-the-art performance on benchmark datasets including CIFAR100, miniImageNet, and CUB200.
• The method shows significant improvements in adapting to new classes and resisting catastrophic forgetting.
• Experiments on the large-scale ImageNet ILSVRC2012 dataset validate the efficacy of the Limit paradigm.

Conclusions:

• Limit offers an effective solution for few-shot class-incremental learning by leveraging meta-learning and a novel calibration module.
• The proposed approach successfully balances the recognition of new classes with the retention of old class knowledge.
• Limit provides a scalable and high-performing framework for real-world applications requiring continuous learning from limited data.