Mahalanobis Distance

Follow this topic Ask about this topic

Mahalanobis distance weights differences by feature covariance, making it useful when dimensions vary in scale or are correlated. Researchers use class-conditional statistics, covariance correction and whitening for out-of-distribution and anomaly detection, including in images, industrial inspection and few-shot intent detection; others learn adaptive metrics for fairness or visual reinforcement learning. Recent work also combines these scores with other signals and adapts them to new domains using pretrained embeddings. A central tension is that results depend strongly on representation quality, covariance estimation and evaluation protocols, while non-Gaussian embeddings can make distance-based confidence scores poorly calibrated.

Latest papers

13 papers1 paper in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.

ReMaD: Tuning-free Domain Adaptation for Classification and Out-of-Distribution Detection
Oct 5, 2026Elijah Bolluyt, Cristina ComaniciuPrototype-Based Classification Mahalanobis Distance
Stevens Institute of Technology, Hoboken, NJ, 07030, USA
Whitening Inverts the Hierarchy: What the Norm of a Whitened Embedding Measures
Sep 19, 2026Mohammed Ahnouch, Lotfi ElaachakFeature Whitening OOD Detection
Université Paris 1 Paris, France · Faculty of Science and Technology of Tangier Abdelmalek Essaadi University Tangier, Morocco
Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance
Sep 1, 2026Jayasimha Talur, Oleg Smirnov, Paul MissaultMahalanobis Distance Intent Classification
Amazon
Decodable but Not Accessible: Auditing Distance-Based Reliability Estimation on Disentangled Skin-Lesion Representations
Aug 10, 2026Duc-Vinh TranDomain-Adversarial Learning Disentangled Representation Learning
School of Information and Communication Technology, Hanoi University of Science and Technology, 1 Dai Co Viet Road, Hai Ba Trung, Hanoi 100000, Vietnam
PAMD: Structured Adaptive Distances for Bisimulation Representations in Visual Reinforcement Learning
Jul 20, 2026Daegyeong Roh, Juho Bae, Han-Lim ChoiMahalanobis Distance Metric Learning
Department of Aerospace Engineering, Korea Advanced Institute of Science and Technology, Daejon, Republic of Korea.
What Does Goodness Measure? A Likelihood-Ratio Account of Forward-Forward Learning
Jul 14, 2026Paolo GiannitrapaniMahalanobis Distance Likelihood Ratio Testing
Department of Information Engineering, Electronics and Telecommunications (DIET) Sapienza University of Rome Rome, Italy
VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models
Jun 19, 2026Florian Seligmann, Emiliyan Gospodinov, Enes Ulas Dincer +1Failure Detection for VLA Models Mahalanobis Distance
Autonomous Learning Robots Lab, Karlsruhe Institute of Technology · FZI Forschungszentrum Informatik, Karlsruhe
Comparing Linear Probes with Mahalanobis Cosine Similarity
Jun 17, 2026Zhuofan Josh Ying, Peter Hase, Nikolaus KriegeskorteLinear Probing Mahalanobis Distance
Departments of 1Psychology, 2Neuroscience, 3Electrical Engineering · 4Zuckerman Mind Brain Behavior Institute; Columbia University, New York, NY · 5Stanford University, Stanford, CA +1
Mahalanobis PatchCore: Covariance-Aware and Streaming-Compatible Industrial Anomaly Detection
May 26, 2026Niccolò Ferrari, Oligert Osmani, Evelina LammaMahalanobis Distance Industrial Anomaly Detection
Department of Engineering, University of Ferrara, Via Saragat 1, 44122, Ferrara, Italy · NAIS s.r.l., Via Maria Callas 4, 40131, Bologna, Italy · Bonfiglioli Engineering S.r.l., Via Amerigo Vespucci 20, 44124, Ferrara, Italy
Operationalizing Individual Fairness via Gradient Descent and Bradley-Terry Models
May 22, 2026Conlan Olson, Linjun Zhang, Zhun Deng +1Mahalanobis Distance Algorithmic Fairness
Columbia University · Rutgers University · UNC Chapel Hill +1
MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse
May 14, 2026Donghwan Kim, Hyunsoo YoonMahalanobis Distance Diffusion-Based OOD Detection
Department of Industrial Engineering, Yonsei University, Seoul, South Korea
Toxicity Assessment in Preclinical Histopathology via Class-Aware Mahalanobis Distance for Known and Novel Anomalies
Feb 2, 2026Olga Graf, Dhrupal Patel, Peter Groß +3Histopathology Image Segmentation Mahalanobis Distance
Tübingen AI Center, University of Tübingen, Tübingen, Germany · Toxicologic Pathologist, Gujarat, India · Boehringer Ingelheim Pharma GmbH and Co., Biberach an der Riß, Germany +1
Auditing Frozen-Encoder Anomaly Detection Across Mechanical Systems: Representation Provenance, Calibration, and Protocol Effects
Jan 16, 2026Jose Sánchez AndreuMahalanobis Distance Industrial Anomaly Detection
Independent Researcher, Murcia, Spain

Oct 5, 2026J/K move · Enter open · S save

ReMaD: Tuning-free Domain Adaptation for Classification and Out-of-Distribution Detection

Elijah Bolluyt, Cristina Comaniciu

Stevens Institute of Technology, Hoboken, NJ, 07030, USA

Prototype-Based Classification Mahalanobis Distance Transfer Learning Classification OOD Detection

Open paper AskPDF arXiv

We introduce Reduced-rank Mahalanobis Distance (ReMaD), a novel prototypical distance-based refinement to classification and out-of-distribution (OOD) detection using pretrained models without finetuning. We use embeddings of the target dataset to fit closed-form distribution statistics in the model's latent space which can classify in-distribution samples and detect OOD samples, all without training or prior knowledge of the OOD data. Building on prototype classification and OOD detection, we analyze the distribution properties of large pretrained models when processing new datasets; based on this analysis, we formulate a simple modification to Mahalanobis Distance to adapt models' latent space distributions to new domains by removing unused features, without the finetuning or hyperparameter searches required by other adaptation procedures. We demonstrate the efficacy of this method to adapt existing large pretrained image embedding models to new classification domains outside their trained capabilities by testing across four target datasets, with competitive performance in both classification and OOD detection.