AI Paper Atlas

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

Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors

Tejaswi Kasarla, Max van Spengler, Pascal Mettes

IJCV 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace Euclidean head with balanced hyperbolic prototypes for MSP scoringResNet-34 MSP · CIFAR-100 (OpenOOD, avg over OOD sets)FPR@95: 58.24 -> 49.46Improved
replace Euclidean head with balanced hyperbolic prototypes for SCALE scoringResNet-34 SCALE · CIFAR-100 (OpenOOD, avg over OOD sets)AUROC: 79.68 -> 79.20Mixed
replace Euclidean head with balanced hyperbolic prototypes for KNN scoringKNN · ImageNet100 (OpenOOD, avg over OOD sets)FPR@95: 37.13 -> 45.74Hurt
use hyperbolic one-hot prototypes instead of Euclidean classifier (ablation)Euclidean baseline · CIFAR-100 (5 scoring functions)FPR@95 / AUROC: 4/5 (FPR@95) and 3/5 (AUROC) scoring functionsImproved
add hierarchical distortion objective and subhierarchy balancing on hyperbolic one-hot prototypes (ablation)Hyperbolic embeddings · CIFAR-100 (5 scoring functions)OOD metricsImproved
add norm loss to distortion loss (ablation)Balanced Hyperbolic Embeddings · toy imbalanced treenorm varianceImproved
use curvature c=1 instead of 0.5 (ablation)Balanced Hyperbolic · CIFAR-100 curvature ablation (MSP)FPR@95: 73.40 -> 49.46Improved
use curvature c=1 instead of 2.0 (ablation)Balanced Hyperbolic · CIFAR-100 curvature ablation (MSP)FPR@95: 85.29 -> 49.46Improved
vary embedding dimension from 8 to 64 (ablation)Balanced Hyperbolic · CIFAR-100MAP: 0.84 -> 0.88No clear effect
train with hierarchy via hyperbolic prototypes instead of Euclidean baselineEuclidean baseline · CIFAR-100 OSR splits (hierarchical OOD, CIFAR-OOD split)H-Dist / HSI-b1: 3.25 -> 2.32 H-Dist; 31.83 -> 67.21 HSI-b1Improved
balanced hyperbolic embeddings instead of Poincare EmbeddingsMSP scoring · CIFAR-100 (MSP scoring, vs other hyperbolic)FPR@95 / AUROC: 50.50 -> 49.46 FPR@95; 83.48 -> 82.43 AUROCMixed
balanced hierarchical hyperbolic prototypes instead of hyperspherical PALM prototypesKNN · CIFAR-100 (KNN k=300, 128-dim prototypes)FPR@95 / n-AUROC: 38.27 -> 35.83 FPR@95; 78.96 -> 78.50 n-AUROCImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy