Learn by Reasoning: Analogical Weight Generation for Few-Shot Class-Incremental Learning

2025-03-27

AI Summary

To address the challenges in few-shot class-incremental learning (FSCIL)—namely, extremely limited samples for novel classes, the need for parameter fine-tuning, and catastrophic forgetting of old knowledge—this paper proposes BiAG, a brain-inspired analogical weight generation framework that requires no parameter updates. Methodologically, BiAG introduces: (1) an analogical generator that formalizes human analogical reasoning as cross-class semantic mapping; (2) Weight Self-Attention (WSA) and Weight & Prototype Analogical Attention (WPAA), enabling zero-shot, fine-tuning-free generation of novel-class classifiers; and (3) a Semantic Conversion Module (SCM) guided by neural collapse theory to enhance feature discriminability and generalization. Evaluated on miniImageNet, CUB-200, and CIFAR-100, BiAG significantly outperforms state-of-the-art methods, achieving substantial improvements in both final accuracy and average incremental accuracy.

Technology Category

• Cognitive Modeling & Cognitive Systems: Analogy
• Machine Learning: Bio-inspired Learning
• Multiagent Systems: Multiagent Learning

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
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Few-shot class-incremental Learning (FSCIL) enables models to learn new classes from limited data while retaining performance on previously learned classes. Traditional FSCIL methods often require fine-tuning parameters with limited new class data and suffer from a separation between learning new classes and utilizing old knowledge. Inspired by the analogical learning mechanisms of the human brain, we propose a novel analogical generative method. Our approach includes the Brain-Inspired Analogical Generator (BiAG), which derives new class weights from existing classes without parameter fine-tuning during incremental stages. BiAG consists of three components: Weight Self-Attention Module (WSA), Weight&Prototype Analogical Attention Module (WPAA), and Semantic Conversion Module (SCM). SCM uses Neural Collapse theory for semantic conversion, WSA supplements new class weights, and WPAA computes analogies to generate new class weights. Experiments on miniImageNet, CUB-200, and CIFAR-100 datasets demonstrate that our method achieves higher final and average accuracy compared to SOTA methods.

Problem

• FSCIL learns new classes with limited data while preserving old class performance
• Traditional FSCIL struggles with knowledge integration and parameter fine-tuning
• Proposes brain-inspired analogical generation for weight creation without fine-tuning

Innovation

• Brain-Inspired Analogical Generator (BiAG) for weight generation
• No parameter fine-tuning during incremental learning
• Neural Collapse theory for semantic conversion