Research map / Label-efficient and robust learning / OOD and anomaly detection
Gradient-Based Novelty Detection Boosted by Self-Supervised Binary Classification
Jingbo Sun, Li Yang, Jiaxin Zhang, Frank Liu, Mahantesh Halappanavar, Deliang Fan, Yu Cao
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 |
|---|---|---|---|
| select an OOD gradient label to maximize Mahalanobis distance rather than use the predicted label (ablation) | Gradient-based novelty detector · CIFAR-10 IDD / TinyImageNet, SVHN, LSUN, and CIFAR-100 OOD | AUROC: 91.10 -> 99.92; 91.63 -> 99.99; 90.08 -> 99.99; 86.03 -> 97.99 | Improved |
| add the self-supervised binary classifier to guide label selection (ablation) | Gradient-based novelty detector · Four OOD test cases in the ablation study | AUROC | Improved |
| train the binary classifier on detector-selected high- and low-score samples instead of randomly selected samples (ablation) | Self-supervised binary classifier · OOD detection experiments | AUROC | Improved |
| use fewer detector-selected IDD/OOD images to train the binary classifier (ablation) | Self-supervised binary classifier · OOD detection experiments | AUROC | Hurt |
| re-initialize the binary classifier after each iteration (ablation) | Self-supervised binary classifier · CIFAR-10 IDD / CIFAR-100 OOD | AUROC | Improved |
| increase binary-classifier batch size from 8 to 32 or 128 (ablation) | Proposed detector · CIFAR-10 IDD / CIFAR-100 OOD | AUROC: 84.79 -> 93.51 -> 97.99 | Improved |
| increase binary-classifier batch size from 8 to 32 or 128 (ablation) | Proposed detector · CIFAR-100 IDD / CIFAR-10 OOD | AUROC: 72.61 -> 89.76 -> 91.38 | Improved |
| increase binary-classifier batch size (ablation) | Proposed detector · Multiple IDD/OOD pairs | AUROC | Mixed |
| use last-layer feature-extractor gradients with Mahalanobis distance instead of only feedforward features | Proposed novelty detector · OOD detection | AUROC | Improved |
| train the binary classifier from detector-generated pseudo-labels rather than labeled OOD data | Proposed novelty detector · OOD detection | AUROC | Improved |
| use the proposed detector for one-class continual learning with a memory budget of 1,000 | CIFAR-10 9+1 single-head learning · CIFAR-10 new-class recognition | accuracy: 91-96% accuracy in any class | Improved |
Closest papers by meaning
- Generalized Out-of-Distribution Detection: A Survey (IJCV 2024 · similarity 0.87)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.86)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.85)
- Universal Novelty Detection Through Adaptive Contrastive Learning (CVPR 2024 · similarity 0.85)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.85)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.85)
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.
