CANNE: CLIP-Based ANNE Selection for Noisy-Label Learning
Ge Jin, Qian Zhang, Li Huang, Yongqi Yuan, Yulong Qiao, De Yu, Mengjie Ye, Yingwen Zhu · 2026
Read CANNE: CLIP-Based ANNE Selection for Noisy-Label Learning by Ge Jin, Qian Zhang, Li Huang, Yongqi Yuan, Yulong Qiao, De Yu, Mengjie Ye, Yingwen Zhu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13512089.txt.
Description
Jin and colleagues published a 2026 study on teaching computers with mislabeled training photos. Wrong labels poison image models, and early mistakes tend to snowball. Their method, CANNE, first borrows a vision-language model's judgment to seed clean examples. It then refines the set during training using neighbor patterns in feature space.
On a standard photo set with 90 percent wrong labels, accuracy stayed above 96 percent. A harder hundred-class set also held up well under heavy noise. Real-world noisy sets from animals and the web confirmed the pattern. Repeated runs showed the seed set steadied training from the start. External priors, they argue, beat self-judgment for early cleaning. Open-set noise, from classes outside training, also tested the framework.
Catalog record
| CREATOR | Ge Jin, Qian Zhang, Li Huang, Yongqi Yuan, Yulong Qiao, De Yu, Mengjie Ye, Yingwen Zhu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 70 KB |
| LICENSE | https://creativecommons.org/licenses/by/4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |