跳转至

文章背景与核心概要

在深度学习模型的训练过程中，理解并预测最终的泛化性能一直是核心挑战之一。传统的验证方法通常依赖于完整的训练周期或昂贵的评估计算。本文作者 Arunan J. 通过实证研究探讨了在 ResNet 微调过程中观察到的离散类可分性跃迁（即“相变”）频率，并评估了其作为预测最终测试准确率指标的有效性。

研究通过横跨多个标准基准、架构和随机种子的大量实验，揭示了标准独立同分布（i.i.d.）基准上相变次数与最终性能之间存在显著的强负相关性。尽管在面对分布外压力（distributional stress）时这种预测能力有所减弱，但该指标作为一种计算成本极低、可直接嵌入标准训练循环的训练质量探测手段，仍展现出极高的实用价值和研究潜力。

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

arXiv: 2609.05194 [cs.LG]
Author: Arunan J
Submitted: September 4, 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)

arXiv: 2609.05194 [cs.LG]
Author: Arunan J
Submitted: September 4, 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)

📌 Summary

本文从实证角度研究了在 ResNet 微调过程中观察到的离散类可分性跃迁（相变）频率，是否能够有效预测最终的测试准确率。通过对多个标准基准、架构和随机种子的 75 组实验进行分析，研究表明在标准的独立同分布（i.i.d.）基准上存在强负相关性，尽管这种预测能力在分布压力下会有所减弱。作者提出将该指标作为一种廉价的训练质量探测器，可以直接在标准训练循环中进行记录。

📌 Summary

This paper empirically investigates whether the frequency of discrete class-separability jumps (phase transitions) observed during the fine-tuning of ResNets can effectively predict final test accuracy. Analyzing 75 experiments across multiple standard benchmarks, architectures, and random seeds, the study demonstrates strong negative correlations on standard i.i.d. benchmarks, though this predictive power attenuates under distributional stress. The author proposes this metric as an inexpensive training-quality probe that can be logged directly alongside a standard training loop.

📝 Abstract

本文实证考察了 ResNet 微调过程中观察到的离散类可分性跃迁次数，并将其作为预测最终测试准确率的指标。研究涵盖了四个基准（CIFAR-10、CIFAR-100、TinyImageNet 和 CIFAR-10-C）和三种架构（ResNet-18、ResNet-50 和 ResNet-101），每种配置使用 5 到 10 个随机种子，共计 75 组实验。在标准的独立同分布（i.i.d.）分类基准上获得了极强的组内负相关性：CIFAR-10 上 \(r = -0.84\)（\(p < 10^{-8}\)，\(n = 30\)），CIFAR-100 上 \(r = -0.87\)（\(p < 10^{-5}\)，\(n = 15\)）。

在分布压力下，这种关系有所减弱：TinyImageNet 的结果为 \(r = -0.45\)，CIFAR-10-C 损坏基准的结果为 \(r = -0.19\)。为了严格论证这一实证主张，进行了两项额外的分析： * 偏相关性（Partial Correlation）： 将架构深度作为线性协变量进行控制，结果表明在 CIFAR-100 上，跃迁计数仍保留了统计学上显著的预测能力（\(r_{\mathrm{partial}} = -0.69\)，\(p = 0.007\)）。在更严格的分类条件设定下，相应的结论在 \(n = 15\) 时尚未确立。 * 基线比较（Baseline Comparison）： 与六种替代的训练曲线信号相比，跃迁计数在 CIFAR-100 上达到了所评估信号中最强的相关性，在 CIFAR-10 上也是最强的信号之一，但在两个受压力的基准上被其他信号超越。

注：此比较仅限于训练曲线级别的信号；与当前文献中最强竞争对手（如有效秩、Hessian 锐度、费雪信息、边界（margin）以及神经崩塌（neural-collapse）度量）的比较超出了本研究的范围，仍有待未来探讨。

该观测结果被提出作为候选探测器家族中的一种分布内训练质量探测手段，并配备了一种适用于在标准训练循环中进行日志记录的低成本检测程序。

📝 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: * Partial Correlation: Controlling for architecture depth 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 stricter categorical conditioning is not established at \(n = 15\). * Baseline Comparison: 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, though it is dominated by other signals on the two stressed benchmarks.

Note: The comparison is restricted to training-curve-level signals; comparisons against effective rank, Hessian sharpness, Fisher information, margin, and neural-collapse measures—which represent the strongest competitors in current literature—are outside the scope of this study and remain open.

The observation is presented as an in-distribution training-quality probe among a family of candidate probes, complete with an inexpensive detection procedure suitable for logging alongside a standard training loop.

🔗 Access & Resources

🔗 Access & Resources

📚 References & Citation Tools

📚 References & Citation Tools