Research map / Continual learning / Class-incremental
Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping
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 GDDSG, a graph-coloring-based dynamic grouping method for class-incremental learning that groups dissimilar classes into isolated sub-models with a frozen ViT backbone, aiming for robustness to class order.
Problem
Class incremental learning performance is highly sensitive to class arrival order, especially when classes arriving together are highly similar (intra-task conflicts).
Approach
Theoretically shows lower inter-class similarity reduces order sensitivity, then builds similarity graphs among class centroids, colors them (Welsh-Powell) to form groups of dissimilar classes, trains a separate NCM/least-squares sub-model per group, and uses a meta-feature classifier to route test samples to groups.
Key results
- GDDSG reaches 94.00% average accuracy on Split CIFAR100 with 0.78% forgetting, versus 90.50% / 3.49% for RanPAC.
- On CUB200, GDDSG achieves 91.64% AN and 1.92% FN; on Dog 92.64% AN and on OB 87.33% AN, outperforming all listed baselines.
- Ablation shows removing class groups drops CIFAR100 AN from 93.99 to 74.32.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDDSG instead of RanPAC | CIL with ViT-B/16 · CIFAR100 | AN: 90.50 ± 0.34 -> 94.00 ± 0.03 | Helped |
| Use GDDSG instead of RanPAC | CIL with ViT-B/16 · CUB200 | AN: 89.23 ± 0.36 -> 91.64 ± 1.86 | Helped |
| Use GDDSG instead of RanPAC | CIL with ViT-B/16 · Dog | AN: 85.37 ± 0.41 -> 92.64 ± 0.48 | Helped |
| Use GDDSG instead of RanPAC | CIL with ViT-B/16 · OB | AN: 78.86 ± 0.16 -> 87.33 ± 0.32 | Helped |
| Add Class Groups (ablation) | GDDSG · CIFAR100 | AN: 74.32 -> 93.99 | Helped |
| Add Class Groups (ablation) | GDDSG · CUB200 | AN: 72.86 -> 92.95 | Helped |
| Add Class Groups (ablation) | GDDSG · Dog | AN: 69.49 -> 92.30 | Helped |
| Add Class Groups (ablation) | GDDSG · OB | AN: 66.85 -> 87.56 | Helped |
| Add SimGraphs and Class Groups (ablation) | GDDSG · CIFAR100 | AN: 89.96 -> 93.99 | Helped |
| Add SimGraphs and Class Groups (ablation) | GDDSG · CUB200 | AN: 87.32 -> 92.95 | Helped |
| Add SimGraphs and Class Groups (ablation) | GDDSG · Dog | AN: 83.12 -> 92.30 | Helped |
| Add SimGraphs and Class Groups (ablation) | GDDSG · OB | AN: 74.21 -> 87.56 | Helped |
Surprising findings
- Smaller ResNet backbones increase the number of class groups for CIFAR100 and CUB200 but decrease it for Stanford Dogs and OB, indicating backbone effects are dataset-dependent.
Limitations
- Relies on NCM classifiers; does not explore softmax-based classifiers.
- Memory overhead, while small, could be further reduced.
- Class-group identification accuracy directly determines overall accuracy.
Experimental setup
Datasets: CIFAR100 CUB200 Stanford Dogs OmniBenchmark
Benchmarks: CIFAR100 CUB200 Stanford Dogs OmniBenchmark
Metrics: average final accuracy average forgetting MOPD AOPD
Baselines: Finetune L2P DualPrompt CODA-Prompt SimpleCIL ADAM EASE RanPAC APD APDfix HALRP
Builds on: RanPAC APD HALRP Welsh-Powell graph coloring Brooks' theorem
Closest papers by meaning
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.86)
- Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach (NeurIPS 2024 · similarity 0.86)
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.86)
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · similarity 0.86)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.