Articles | Volume 33, issue 4
https://doi.org/10.5194/npg-33-503-2026
https://doi.org/10.5194/npg-33-503-2026
Research article
 | 
02 Oct 2026
Research article |  | 02 Oct 2026

Conditional updates of neural network weights for increased out of training performance

Jan Saynisch-Wagner and Saran Rajendran Sari

Viewed

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 1,573 (including HTML, PDF, and XML)
HTML PDF XML Total BibTeX EndNote
1,371 137 65 1,573 60 56
  • HTML: 1,371
  • PDF: 137
  • XML: 65
  • Total: 1,573
  • BibTeX: 60
  • EndNote: 56
Views and downloads (calculated since 23 Feb 2026)
Cumulative views and downloads (calculated since 23 Feb 2026)

Viewed (geographical distribution)

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 1,573 (including HTML, PDF, and XML) Thereof 1,558 with geography defined and 15 with unknown origin.
Country # Views %
  • 1
1
 
 
 
 
Latest update: 17 Nov 2026
Download
Short summary
Neural networks are limited in situations that differ from the learned conditions. We propose a 3-step solution to this out of distribution problem: (1) Derive the anomalies of a trained neural networks' internal parameters by retraining on subsets of its training data. (2) Relate the ensuing network-parameter anomalies to differences in the training data subsets that caused them. (3) Extrapolate the found relations to generate a network that performs better outside the org. training distribution.
Share