Research map / Continual learning / Class-incremental

Attraction Diminishing and Distributing for Few-Shot Class-Incremental Learning

CVPR 2025 · Continual learning › Class-incremental · approach: Few-shot class-incremental prototypes · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Analyzes hubness in nearest-neighbor classification for few-shot class-incremental learning and proposes D2A, which diminishes hub attraction via a Manhattan-blended distance metric at inference and distributes it via a gap-sample loss during base training.

Problem

Few-shot class-incremental learning relies on nearest-neighbor prototype classification, which suffers from hubness: base-class prototypes become hubs that attract novel-class test samples, hurting incremental-class accuracy.

Approach

Theoretically analyzes hubness under neural collapse and cosine distance, then (1) blends cosine distance with Manhattan distance at inference with an adaptive weight driven by MMD between base and incremental prototypes, and (2) adds a gap-sample loss on midpoints of nearest cross-class samples during base training to create more component distributions.

Key results

• Applying D2A to SAVC raises last-session overall accuracy on miniImageNet from 55.97% to 57.72%, and its incremental class accuracy in session 1 from 38.00% to 46.00%.
• Applying D2A to CEC raises last-session overall accuracy on miniImageNet from 47.63% to 51.34%.
• Applying D2A to Decoupled-Cosine raises last-session overall accuracy on miniImageNet from 45.92% to 49.33%.

Measured effects reported in the paper

Surprising findings

• Chebyshev distance gives the largest incremental-class gain but distorts similarity relations enough to hurt overall accuracy, so the metric choice involves a trade-off.

Limitations

• Distributing attraction is not applied to CLOM and SAVC on CUB200 due to fine-grained feature complexity, so gains there are not shown.
• Hyperparameters (eta, alpha, r) are set per dataset and baseline, requiring tuning.
• Theoretical analysis relies on assumptions such as i.i.d. coordinates and neural collapse holding approximately.

Experimental setup

Datasets: miniImageNet CIFAR100 CUB200

Benchmarks: miniImageNet CIFAR100 CUB200

Metrics: incremental class accuracy overall accuracy average incremental class accuracy

Baselines: FACT C-FSCIL TEEN Bidist OSHHG EHS DyCR CLOSER M2SD Decoupled-Cosine CEC CLOM

Builds on: Decoupled-Cosine CEC CLOM SAVC neural collapse

Closest papers by meaning

• M2SD:Multiple Mixing Self-Distillation for Few-Shot Class-Incremental Learning (AAAI 2024 · similarity 0.83)
• Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.83)
• SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · similarity 0.82)
• Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need (CVPR 2025 · similarity 0.82)
• Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.82)
• OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024 · similarity 0.82)