Deep networks exhibit a phenomenon called 'agreement-on-the-line' under distribution shifts, where in-distribution versus out-of-distribution accuracy is strongly linearly correlated across architectures and hyperparameters, and the agreement between predictions of independently trained networks follows the same linear trend.

Authors

• Person: Samuel Tesfazgi, Leonhard Sprandl, Sandra Hirche Organization: AISTATS
  Track: Poster Session 3 - aistats 2026

Sources

• Track: Poster Session 3 - aistats 2026 virtual.aistats.org Samuel Tesfazgi, Leonhard Sprandl, Sandra Hirche · AISTATS via serper

Referenced by nodes (1)

• deep neural networks concept