Multi-Metric Prototypical Networks for Few-Shot Class-Incremental Learning

Abstract

Few-shot class-incremental learning (FSCIL) requires a model to absorb new visual classes from a handful of labeled examples while preserving decisions for previously seen classes. Prototype insertion is attractive because it avoids incremental optimization, yet a single embedding geometry can make few-shot prototypes brittle. We propose a Multi-Metric Prototypical Network (MMPN) that trains independent cosine, normalized Euclidean, and Mahalanobis branches and combines their logits with learnable simplex weights. After base training, class means replace the linear classifiers. The Mahalanobis branch estimates a shrinkage-regularized full covariance from normalized base embeddings and updates this state with class-count-weighted session statistics. MMPN obtains 65.80% average accuracy on miniImageNet and 66.43% on CIFAR100, exceeding the strongest listed baselines by 4.19% and 3.33%. Under the 11-session CUB200 ablation protocol, it improves average and final-session accuracy over a cosine-only baseline by 1.48% and 1.59%. These results indicate that metric-specific representations and covariance-aware matching provide complementary evidence, at the cost of using three backbone networks.