Research map / Continual learning / Class-incremental

SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning

CVPR 2025 · Continual learning › Class-incremental · approach: Few-shot class-incremental prototypes · 5 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.

Proposes SEC-Prompt, a prompt-based few-shot class-incremental learning framework that splits prompts into discriminative (attention-based) and non-discriminative (reverse-attention) components, with a prompt clustering loss and non-discriminative-prompt data augmentation.

Problem

Few-shot class-incremental learning requires learning new classes from very few samples while avoiding catastrophic forgetting; existing prompt-based CIL methods rely on abundant data and ignore semantic differences between prompt features, causing a plasticity-stability dilemma.

Approach

Model discriminative and non-discriminative features with complementary prompts: D-Prompt uses softmax attention to amplify class-specific features, while ND-Prompt uses reverse attention with ReLU to balance non-discriminative features; ND-Prompt perturbations serve as data augmentation, and a Prompt Clustering Loss regularizes D-Prompt. Queries are the [class] token from each layer's input (adap

Key results

• On ImageNet-R, SEC-Prompt reaches final-session accuracy 74.5% with PD 8.3% and HAcc 74.2%, versus ASP at 69.7% final accuracy, 13.5% PD, and 67.0% HAcc.
• On CUB200, SEC-Prompt achieves PD 2.6% and HAcc 84.7%, and on CIFAR100 PD 4.5% and HAcc 87.1%, best among compared methods on PD and HAcc.
• Ablation on ImageNet-R: replacing Coda-Prompt with D-Prompt plus adaptive hierarchical query raises HAcc from 61.7 to 72.8; full model reaches 74.2.

Measured effects reported in the paper

Surprising findings

• Adding a contrastive loss with negative samples degraded performance, which the authors attribute to shared discriminative traits across categories (e.g., horses and zebras).
• Increasing prompt pool size beyond a peak hurt performance because larger pools amplify forgetting of old knowledge.

Limitations

• Performance on the base session is lower than SOTA FSCIL methods that fine-tune the ViT backbone.
• Larger prompt pools cause more forgetting; the quality of few-shot-learned knowledge remains suboptimal.
• Contrastive negatives reduced performance, so the clustering loss uses only positive pairs.

Experimental setup

Datasets: CIFAR100 CUB200-2011 ImageNet-R

Benchmarks: CIFAR100 CUB200-2011 ImageNet-R

Metrics: Top-1 accuracy per session performance drop (PD) harmonic accuracy (HAcc)

Baselines: TEEN PriViLege ASP L2P DualP CodaP

Builds on: CODA-Prompt L2P Visual Prompt Tuning ASP

Closest papers by meaning

• PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback Regulation for Class-Incremental Learning (AAAI 2025 · similarity 0.89)
• Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.88)
• PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.88)
• Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.88)
• OSLoPrompt: Bridging Low-Supervision Challenges and Open-Set Domain Generalization in CLIP (CVPR 2025 · similarity 0.88)
• Convolutional Prompting meets Language Models for Continual Learning (CVPR 2024 · similarity 0.88)