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

TL;DR

This work proposes the Brain-Inspired Analogical Generator (BiAG), which derives new class weights from existing classes without parameter fine-tuning during incremental stages, and achieves higher final and average accuracy compared to SOTA methods.

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.

Figures (5)

• Figure 1: An example of human analogical learning mechanism and the proposed Brain-Inspired Analogical Generator (BiAG). (a) Humans can efficiently form concepts of new categories through analogical reasoning. For example, when encountering a new animal like a panda, one may relate its visual characteristics to previously known categories—such as the body shape of a bear and the black-and-white color pattern of a zebra. This analogy enables the construction of a concept of "panda" even from limited samples. (b) Inspired by this cognitive mechanism, our BiAG module analogically generates classification weights for new classes by leveraging prototypes and weights from old classes along with a learnable query derived from new class knowledge. This allows efficient integration of new categories in few-shot incremental learning.
• Figure 2: An overview of the proposed Analogical Generative Method.Knowledge Base Construction: The network is trained to obtain classification weights and prototypes for the base classes. Analogical Generator Training: The BiAG uses pseudo-incremental training and ${\mathcal{L}}_{G}$ to ensure that generated weights match true weights. Weight Generation: In the t-th session, BiAG generates weights for the new classes using new class prototypes, old class prototypes, and old class weights.
• Figure 3: The architecture of the proposed Brain-Inspired Analogical Generator (BiAG). BiAG consists of stacked layers of Weight Self-Attention and Weight & Proto-Analogical Cross-Attention modules. A shared Semantic Conversion Module (SCM) transforms learnable queries between prototype and weight semantics to facilitate analogical generation of new class weights.
• Figure 4: The results of ablation experiments.Ablation studies on the proposed BiAG. (a) Final session accuracy on three datasets after incrementally adding key components (WPAA, WSA, SMC); (b) Session-wise accuracy comparison on CUB-200, showing the contribution of each component throughout the incremental process; (c) Accuracy comparison of BiAG with different numbers of layers, demonstrating that a deeper design improves generalization to new classes.
• Figure 5: t-SNE visualization on CIFAR-100. (a) Base session: clear class clusters aligned with classifier weights. (b) Incremental session: BiAG-generated weights (red) better match new class features than prototype-based weights (black), preserving structure and improving generalization.