AuraTracer智迹闻
中文

EVENT DOSSIER

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

2026-09-07 12:00 Science across 2 days 🔥 47.2 heat score
2sources
2days unfolding
47.2heat score
5mentions
SummaryAI generated

A study on the fine-tuning process of ResNet found a significant strong negative correlation between the number of discrete category separability transitions and the final test accuracy. In 75 experiments covering four benchmark datasets and three network architectures, the CIFAR-10 and CIFAR-100 standard independent and uniform distribution classification tasks showed correlation coefficients of -0.84 and -0.87, respectively (p < 10^-8 and p < 10^-5). Under TinyImageNet and CIFAR-10-C distribution stress tests, this correlation weakened to -0.45 and -0.19. After controlling for network depth, the partial correlation coefficient on CIFAR-100 remained at -0.69 (p = 0.007). Compared with six alternative training curve signals, the number of transitions performed well in predicting test accuracy.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
CIFAR-10CIFAR-100ResNetTinyImageNetarXiv

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
CIFAR-10 × CIFAR-1001CIFAR-10 × ResNet1CIFAR-10 × TinyImageNet1CIFAR-100 × ResNet1CIFAR-100 × TinyImageNet1ResNet × TinyImageNet1

Integrated timelineUNIFIED TIMELINE

  1. 2026-09-04

    Phase Transition Frequency as a Trainin…

    研究人员通过 75 项实验,实证检验离散类可分性跳跃数量作为 ResNet 测试准确率训练时间预测器的有效性。在 CIFAR-10 和 CIFAR-100 基准上分别获得强负相关系数(r = -0.84 和 r = -0.87),而在 T…

  2. 2026-09-07

    Phase Transition Frequency as a Trainin…

    一项针对 ResNet 微调过程的研究发现,离散类别可分性跳变数量与最终测试准确率存在强负相关。在涵盖四个基准数据集和三种网络架构的 75 次实验中,CIFAR-10 和 CIFAR-100 标准独立同分布分类任务分别呈现 \(r = -…

SignalsSIGNALS

Keyword heat
  • arXiv1
  • ResNet1
  • CIFAR-101
  • CIFAR-1001
  • TinyImageNet1

All reports (2)SOURCES

A arXiv cs.LG en 2026-09-04 22:28

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

研究人员通过 75 项实验,实证检验离散类可分性跳跃数量作为 ResNet 测试准确率训练时间预测器的有效性。在 CIFAR-10 和 CIFAR-100 基准上分别获得强负相关系数(r = -0.84 和 r = -0.87),而在 TinyImageNet 和 CIFAR-10-C 分布压力基准下该关系减弱。控制网络深度后,CIFAR-100 上的部分相关系数仍具统计显著性(r_partial = -0.69)。对比六种替代训练曲线信号发现,跳跃计数在 CIFAR-100 上相关性最强,在 CIFAR-10 上位列第二,但在压力基准下被其他信号主导。本研究未包含有效秩、海森矩阵尖锐度等外部竞争指标的比较。该观察被视为一种分布内训练质量探针,并提供了一种适合与标准训练循环并行的低成本检测流程。

A arXiv cs.AI en 2026-09-07 12:00

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

一项针对 ResNet 微调过程的研究发现,离散类别可分性跳变数量与最终测试准确率存在强负相关。在涵盖四个基准数据集和三种网络架构的 75 次实验中,CIFAR-10 和 CIFAR-100 标准独立同分布分类任务分别呈现 \(r = -0.84\) 和 \(r = -0.87\) 的相关系数(\(p < 10^{-8}\) 和 \(p < 10^{-5}\))。在 TinyImageNet 和 CIFAR-10-C 分布压力测试下,该相关性减弱至 \(r = -0.45\) 和 \(r = -0.19\)。控制网络深度后,CIFAR-100 上的部分相关系数仍达 \(r_{\mathrm{partial}} = -0.69\)(\(p = 0.007\))。对比六种替代训练曲线信号显示,跳变计数在 CIFA…