Anomaly-ranked confirmation slows new-class learning
Abstract
Anomaly detectors rank unfamiliar examples for expert confirmation, and the confirmed examples teach a classifier to recognise a new class. Across five benchmarks from physics, medicine and computer vision, the members flagged by the detector are those the classifier learns first. In three systems, confirming the highest-ranked members of a new class gives lower recognition than confirming the same number of random members. Handing subsequent selection to the promoted classifier recovers part of this loss in two systems. A paired intervention further shows that replacing nearby known classes with distant ones sharply reduces the confirmations needed for recognition. These results distinguish finding a new class from selecting examples that teach its extent: within the tested classes, concentrating confirmations at the top of the anomaly ranking makes learning less efficient.
Related
- Rotation slows the instability of Ellis-Bronnikov wormholes
- Bekenstein--Hod Bound: A 3.3σ Confirmation from GW250114
- EMIR confirmation of the high-z Lyα blob SHARDS20018464
- Searching for Embedded Protoplanets with the Keck/NIRC2 Vortex Coronagraph: Confirmation of a Core-accretion Planet in the Narrow Gap of the Elias 2-24 Disk
- EMIR confirmation of the high- z Ly α blob SHARDS20018464
BibTeX
@misc{sasli2026anomalyranked,
title = {{Anomaly-ranked confirmation slows new-class learning}},
author = {Argyro Sasli and Michael Coughlin},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.03250},
eprint = {2611.03250},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.03250},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Anomaly-ranked confirmation slows new-class learning
AU - Argyro Sasli
AU - Michael Coughlin
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.03250
UR - https://arxiv.org/abs/2611.03250
AB - Anomaly detectors rank unfamiliar examples for expert confirmation, and the confirmed examples teach a classifier to recognise a new class. Across five benchmarks from physics, medicine and computer vision, the members flagged by the detector are those the classifier learns first. In three systems, confirming the highest-ranked members of a new class gives lower recognition than confirming the same number of random members. Handing subsequent selection to the promoted classifier recovers part of this loss in two systems. A paired intervention further shows that replacing nearby known classes with distant ones sharply reduces the confirmations needed for recognition. These results distinguish finding a new class from selecting examples that teach its extent: within the tested classes, concentrating confirmations at the top of the anomaly ranking makes learning less efficient.
ER -