Research map / Continual learning / Class-incremental
Geometry-driven OOD Detectors Are Class-Incremental Learners
Wangwang Jia, Zijian Gao, Tianjiao Wan, Yuan Cao, Yong Dou, Kele Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use refined top-k sparse task activation instead of coarse single-pass inference (ablation) | GOD · ImageNet-R (T=5) | average incremental accuracy: 85.31 -> 86.90 | Improved |
| use refined top-k sparse activation instead of activating all task LoRA blocks (ablation) | GOD · ImageNet-R (T=5) | average incremental accuracy: 86.74 -> 86.90 | No clear effect |
| use refined mode instead of coarse mode (ablation) | GOD · ImageNet-A (T=5) | inference time (s): 3.79 -> 9.24 | Hurt |
| vary top-k from 2 to 5 (ablation) | GOD (Refined) · ImageNet-R (T=10) | average incremental accuracy: 84.85 -> 85.41 (k=2 to 3), 85.16 at k=5 | No clear effect |
| replace learnable linear/cosine head with fixed ETF anchors plus ArcFace (ablation) | GOD · Stanford Cars / ImageNet-R (two tasks) | task-ID confidence gap | Improved |
| add ArcFace loss on top of ETF loss (ablation) | GOD · Stanford Cars (t-SNE, qualitative) | cluster compactness | Improved |
| add random projection on top of ETF loss (ablation) | GOD · Stanford Cars (t-SNE, qualitative) | global class separation | Improved |
| combine ETF loss, ArcFace and random projection (ablation) | GOD · Stanford Cars (t-SNE, qualitative) | cluster quality | Improved |
| share shallow 9 LoRA blocks across tasks and use task-specific deep 3 blocks | GOD · ImageNet-A (T=5) | trainable parameters (M): 4.73 -> 2.64 | Improved |
| GOD versus strongest baseline on CIFAR-100 | GOD (Refined) · CIFAR-100 (T=10) | average incremental accuracy: 92.01 -> 92.87 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SD-LoRA — SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.86)
- CADRef: Robust Out-of-Distribution Detection via Class-Aware Decoupled Relative Feature Leveraging (CVPR 2025 · similarity 0.86)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.86)
- Learning with Mixture of Prototypes for Out-of-Distribution Detection (ICLR 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.
