Xinchao Wang
National University of Singapore
Anomaly DetectionFrontier Research Maph-index: 7159 citations
Frontier Research Map
Featured Work
Multimodal Models Beyond Human Supervisionofficial homepage invited talks list — 2025-03-01
Why Now
A justified rising-voice inclusion because it sharpens the argument that future visual intelligence will not be bottlenecked by manual human labels.
Key Ideas
- -Self-supervised and multimodal objectives can learn visual concepts beyond the taxonomy of benchmark labels.
- -Language-space features and visual lexicons may let models reuse structure more flexibly across tasks.
- -The next phase of vision may be less about class sets and more about transferable latent interfaces.
Open Questions
- ?What does supervision beyond human labels look like in practice: video, language, interaction, or synthetic structure?
- ?How do we evaluate concepts that have no clean benchmark ontology?
- ?Can open-ended pretraining stay scientifically interpretable?
Younger Agenda-Setters and Adjacent ML Thinkersmedium confidence
Anomaly Detection Research
Research Theme
foundation models and visual outliers
unreviewedlow confidence
Sign in to access this content