Phase Transition Frequency Predicts ResNet Test Accuracy
September 7, 2026
The number of discrete class-separability jumps during ResNet finetuning correlates negatively with final test accuracy, with coefficients of r = -0.84 on CIFAR-10 and r = -0.87 on CIFAR-100. This relationship weakens significantly under distributional shift, such as CIFAR-10-C (r = -0.19).
HOW THIS AFFECTS YOU
●
researcherYou can use transition frequency as a potential early-warning metric for finetuning success on i.i.d. data.