Answers are grounded in this paper and cite their sources.
Shows that confusion matrix diagonal strength predicts transfer learning outcomes, indicating its utility in model design.
We present evidence that the diagonal strength of a zero-shot confusion matrix, computed from a single zero-shot evaluation on a labeled target set before any fine-tuning, predicts transfer learning outcome in two distinct regimes. On structure-preserving geometric shifts such as rotations and blur, higher diagonal strength predicts a more favorable transfer outcome. On structure-destroying degradations such as Gaussian noise, salt-and-pepper noise, and contrast reduction, lower diagonal strength predicts a larger transfer benefit. The same scalar carries opposite predictive information in the two shift regimes. In the rotation experiments, the binary prediction correlation was r = 0.445 (p = 0.000006, n = 96) across eight MLP architecture sizes and replicated on blur transforms across two datasets (r = 0.495 and r = 0.564). In the degradation experiments, the magnitude prediction correlation ranged from r = -0.573 to r = -0.941 across four degradation experiments spanning three degradation types and two datasets. We document boundary conditions where the method did not produce useful predictions and report all negative results alongside positive findings. We also report an exploratory secondary signal in off-diagonal entropy and one financial temporal regime-shift validation on Lending Club data, with evaluation on 170,522 held-out test loans. All experiments use real published datasets with no synthetic data. The methods described in this paper are the subject of U.S. Provisional Patent Application No. 63/920,092 (filed November 18, 2025). No license to implement or commercialize the described methods is granted by this publication. All rights reserved.
No takes yet. Share an insight, caveat, or question.
Shawn Barnicle (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: