Research map / Label-efficient and robust learning / OOD and anomaly detection

Universal Novelty Detection Through Adaptive Contrastive Learning

CVPR 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Uncertainty and misclassification detection · 10 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 UNODE, a contrastive-learning novelty detector with AutoAugOOD, which adaptively selects hard augmentations (KL/JS divergence in t-SNE-projected feature space) to craft negative pairs, aiming for universality across train-set and test-set shifts.

Problem

Novelty detection methods lack universality: they fail to transfer across training datasets, degrade under corrupted test samples, and do not adapt across setups (one-class, unlabeled and labeled multi-class), because their inductive biases are rigid and contrastive learning lacks negative pairs for

Approach

Automatically craft negative pairs by applying hard augmentations (from SimCLR's detrimental set) weighted by the Jensen-Shannon divergence between 1-D t-SNE projections of inlier and augmented feature distributions, then train contrastive plus binary cross-entropy losses and score with similarity plus binary inlier probability.

Key results

• UNODE reaches 92.2% mean AUROC over 10 standard one-class datasets versus 92.5 for CSI and a lower STD of 5.5 versus 14.0 for prior methods (per Table 1 pre-trained/from-scratch comparison).
• On corrupted one-class benchmarks UNODE achieves 84.2% mean AUROC, compared with 84.7 for CSI and 80.8 for GT.
• In the ablation, AutoAugOOD yields 97.3 mean AUROC over MNIST, CIFAR-10 and MVTecAD, above fixed augmentations such as Rotate (83.6) and CutPaste (83.6).

Measured effects reported in the paper

Surprising findings

• Fixed hard augmentations such as rotation can hurt performance on rotation-invariant datasets, while the adaptive selection avoids this.
• Standard contrastive loss alone performs poorly for novelty detection (e.g., 61.0 on MVTecAD in the loss ablation).

Limitations

• Corrupted-benchmark mean AUROC of 84.2% is below CSI's 84.7%, so robustness is not uniformly superior.
• Experiments focus on one-class setting; multi-class extensions are shown on a limited set of CIFAR-based cases.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 MNIST EMNIST-Letters FashionMNIST SVHN MVTecAD FGVC-Aircraft Head CT ISIC 2018 ImageNet-21k ImageNet-1k

Benchmarks: CIFAR-10 CIFAR-100 MNIST EMNIST-Letters FashionMNIST SVHN MVTecAD HeadCT FGVC ISIC2018 CIFAR-10-C CIFAR-100-C

Metrics: AUROC mean AUROC standard deviation across datasets

Baselines: DeepSVDD GT MHROT CSI DN2 MSAD PANDA FITYMI FastFlow PatchCore ReContrast SSD

Builds on: SimCLR CSI AutoAugment PANDA MSAD

Closest papers by meaning

• CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances (NeurIPS 2020 · similarity 0.90)
• AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing (NeurIPS 2025 · similarity 0.85)
• Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.84)
• Novelty Detection in Reinforcement Learning with World Models (ICML 2025 · similarity 0.83)
• One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.83)
• An Interpretable Evaluation of Entropy-based Novelty of Generative Models (ICML 2024 · similarity 0.83)