EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Noise-adaptive cross-modal training with evidential deep learning-based uncertainty

Genna Ba, Haobin Li, Guozi Sun, Ximeng Liu, Kun Guo

Journal of King Saud University Computer and Information Sciences · 2026-11 · 38 (9)

DOI: 10.1007/s44443-026-01344-8 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of King Saud University Computer and Information Sciences, online erschienen am 02.11.2026. Verfasst von Genna Ba und 4 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: The Open Topics from the Lion Rock Labs of Cyberspace Security under the project, National Natural Science Foundation of China under Grant, National Key Research and Development Plan of China under Grant.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Cross-modal retrieval under pairwise noisy supervision is challenging because mismatched image–text pairs are often enforced as positives, distorting the shared embedding space and degrading retrieval robustness. We propose Noise-Adaptive Cross-Modal Training (NACM), a unified framework that combines dynamic reliability estimation with differentiated optimization. NACM improves sample partitioning by fusing global loss statistics with local evidential confidence, where the former captures separability in the loss space and the latter measures model-supported correspondence credibility. Based on this reliability-aware partition, NACM optimizes relatively clean pairs through uncertainty-aware alignment, while exploiting noisy pairs only via uncertainty-gated soft structural regularization. This design avoids both indiscriminate filtering and direct reinforcement of unreliable annotations, allowing clean and noisy samples to play complementary roles during training. Experiments on Flickr30K, MS-COCO, and CC152K demonstrate robust and competitive performance under both synthetic and real-world noisy correspondences. Additional ablation studies and uncertainty-distribution analyses further verify the contributions of the major components and illustrate how estimated uncertainty is translated into differentiated optimization strengths.

Förderung

  • The Open Topics from the Lion Rock Labs of Cyberspace Security under the project
  • National Natural Science Foundation of China under Grant
  • National Key Research and Development Plan of China under Grant
  • Natural Science Foundation of Fujian Province under Grant

Zitieren

APA

Ba, G., Li, H., Sun, G., Liu, X., & Guo, K. (2026). Noise-adaptive cross-modal training with evidential deep learning-based uncertainty. Journal of King Saud University Computer and Information Sciences, 38(9). https://doi.org/10.1007/s44443-026-01344-8

Vancouver

Ba G, Li H, Sun G, Liu X, Guo K. Noise-adaptive cross-modal training with evidential deep learning-based uncertainty. Journal of King Saud University Computer and Information Sciences. 2026;38(9). doi:10.1007/s44443-026-01344-8

BibTeX
@article{Ba2026,
  author = {Ba, Genna and Li, Haobin and Sun, Guozi and Liu, Ximeng and Guo, Kun},
  title = {Noise-adaptive cross-modal training with evidential deep learning-based uncertainty},
  journal = {Journal of King Saud University Computer and Information Sciences},
  year = {2026},
  volume = {38},
  number = {9},
  doi = {10.1007/s44443-026-01344-8}
}
Kostenlos starten →