Life sciences · Preprint

Phase Transition Frequency as a Training Time Predictor of Test Accuracy in ResNets

arXiv · September 4, 2026

Preliminary Why this matters to you Expert discussion

Early or partial results. Treat as a signal, not a conclusion.

AI Summary

This preprint reports a strong negative correlation between the number of discrete class-separability phase transitions during ResNet training and final test accuracy on clean in-distribution benchmarks (CIFAR-10 r = −0.84, CIFAR-100 r = −0.87), but the relationship weakens substantially under distributional stress. The work is descriptive and exploratory, proposing a training-time proxy for model quality; it does not establish causation, generalize beyond vision tasks, or outperform established competitor metrics on stressed data.

What was studied

Empirical correlation study across multiple image classification benchmarks. ResNet models trained on CIFAR-10, CIFAR-100, TinyImageNet, and CIFAR-10-C (corruption variant).. Intervention: Measurement of phase transition frequency as a training-time predictor during standard supervised finetuning.. Compared with: Six alternative training-curve signals; no explicit control condition, as the study is observational and correlational.. n = 75.

Key findings

On CIFAR-10, strong negative correlation between phase transition count and test accuracy: r = −0.84 (p < 10−8, n = 30) On CIFAR-100, strong negative correlation: r = −0.87 (p < 10−5, n = 15) Partial correlation on CIFAR-100 controlling for architecture depth: r_partial = −0.69 (p = 0.007)

Safety

Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.

Who this may help

The source did not state who this applies to in practice.

Plain-language summary

This is an empirical observational study of a training dynamics proxy in neural networks, with correlational evidence across limited benchmarks and no causal inference or clinical translation; the finding is domain-specific to computer vision and does not address a clinical or health outcome.

Study design

As stated by the source record.

MethodsPhase transition frequency (discrete class-separability jumps) was measured during finetuning of ResNet models on four benchmarks. Experiments totalled 75 configurations across three architectures (ResNet-18, 50, 101) with 5–10 random seeds per configuration. Pearson correlations were computed between transition count and final test accuracy. Partial correlation analysis controlled for architecture depth as a linear covariate. Comparison against six alternative training-curve signals was performed on each benchmark. Analysis was restricted to training-curve-level signals; comparisons against…

DesignEmpirical correlation study across multiple image classification benchmarks

PopulationResNet models trained on CIFAR-10, CIFAR-100, TinyImageNet, and CIFAR-10-C (corruption variant).

Sample size75

InterventionMeasurement of phase transition frequency as a training-time predictor during standard supervised finetuning.

ComparatorSix alternative training-curve signals; no explicit control condition, as the study is observational and correlational.

Reported figures

Quoted from the source exactly as published.

Pearson correlation, CIFAR-10r = −0.84

p-value, CIFAR-10p < 10−8

Sample size, CIFAR-1030

Pearson correlation, CIFAR-100r = −0.87

p-value, CIFAR-100p < 10−5

Sample size, CIFAR-10015

Partial correlation, CIFAR-100 (controlling for depth)r_partial = −0.69

p-value, partial correlation CIFAR-100p = 0.007

Caveats

This is a preprint and has not been peer-reviewed
Small sample sizes (n=15 on CIFAR-100); partial correlation analysis underpowered for categorical conditioning tests
Correlation is absent or very weak on out-of-distribution and corruption-stressed data, limiting generalizability
No comparison against several cited competitor signals (effective rank, Hessian sharpness, Fisher information, margin, neural-collapse measures); restricted to training-curve-level signals only

Evidence assessment

Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.

Study design

moderate

Systematic empirical observation across 75 experiments with multiple seeds; no randomization or intervention.

Directness

limited

Measures training dynamics proxy in image classification; remote from clinical or biomedical application.

Consistency

limited

Correlation attenuates markedly under distributional stress; weak on corrupted and out-of-distribution data.

Risk of bias

limited

Unblinded observational work; lacks comparison to most cited competitor signals (Fisher, Hessian, margin, neural collapse).

Sample adequacy

limited

n=30 on CIFAR-10, n=15 on CIFAR-100; small effective samples; partial correlation analysis underpowered at n=15.

Source abstract

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 (\(r_{\mathrm{partial}} = -0.69\), \(p = 0.007\)); 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 count achieved the strongest correlation among the evaluated signals on CIFAR-100 and one of the strongest on CIFAR-10, but is dominated by other signals on the two stressed benchmarks. The comparison is restricted to training-curve-level signals; comparisons against effective rank, Hessian sharpness, Fisher information, margin, and neural-collapse measures, which are the strongest competitors in the current literature, are not part of the present study and remain open. The observation is presented as an in-distribution training-quality probe among a family of candidate probes, and an inexpensive detection procedure suitable for logging alongside a standard training loop is provided.

Follow from this

Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.

Arunan J