Phase Transition Frequency as a Training Time Predictor of Test Accuracy in ResNets
A study examined the number of discrete class-separability jumps during ResNet finetuning as a predictor of final test accuracy. Across 75 experiments on CIFAR-10, CIFAR-100, TinyImageNet, and CIFAR-10-C with ResNet-18, ResNet-50, and ResNet-101, strong negative correlations were found on CIFAR-10 (r = -0.84, p < 10^-8, n = 30) and CIFAR-100 (r = -0.87, p < 10^-5, n = 15). Under distributional stress, correlations weakened: TinyImageNet r = -0.45 and CIFAR-10-C r = -0.19. Partial correlation controlling for architecture depth on CIFAR-100 retained significance (r_partial = -0.69, p = 0.007).
The number of discrete class-separability jumps observed during ResNet finetuning is examined empirically as a predictor of final test accuracy. Across 75 experiments spanning four benchmarks (CIFAR-10, CIFAR-100, TinyImageNet, and CIFAR-10-C) and three architectures (ResNet-18, ResNet-50, and ResNet-101), with five to ten seeds per configuration, a strong within-dataset negative correlation is obtained on standard i.i.d. classification benchmarks: r = -0.84 on CIFAR-10 (p < 10^-8, n = 30) and r = -0.87 on CIFAR-100 (p < 10^-5, n = 15). Under distributional stress, the relationship attenuates: TinyImageNet yields r = -0.45, and the CIFAR-10-C corruption benchmark yields r = -0.19. Two additional analyses discipline the empirical claim. A partial correlation controlling for architecture depth, treated as a linear covariate, shows that on CIFAR-100 the transition count retains statistically significant predictive power (r_partial = -0.69, p = 0.007); the corresponding result under the stricter categorical conditioning is not established at n = 15. A comparison against six alternative training-curve signals shows that transition frequency is competitive with or superior to existing early-stopping heuristics on the i.i.d. benchmarks.
The study introduces phase transition frequency—the count of discrete class-separability jumps during finetuning—as a training-time signal. The strong negative correlation on i.i.d. benchmarks suggests that fewer transitions indicate more stable convergence and higher final accuracy. The attenuation under distribution shift (CIFAR-10-C) implies the signal may be less reliable for robustness prediction. Partial correlation results indicate the effect is not solely due to architecture depth, at least on CIFAR-100.
This research could inform early-stopping and model selection practices in computer vision training pipelines. If transition frequency can be monitored cheaply during training, practitioners might reduce wasted compute on poorly converging runs. However, the weaker correlation on corrupted benchmarks suggests limited applicability for robustness-critical deployments without further validation.
Potential business value lies in reducing training costs by early termination of runs with high transition frequency, and in improving model selection without additional test-time evaluation. However, the current evidence is limited to ResNet finetuning on small-to-medium image classification benchmarks, so immediate commercial impact is uncertain.
Next observable signals include replication on larger-scale datasets and transformer architectures, investigation of the causal mechanism behind phase transitions, and development of practical monitoring tools that compute transition frequency in real time. The authors may also explore whether the signal generalizes to other tasks such as segmentation or detection.