Research map / Label-efficient and robust learning / Recognition and long-tail
Let’s Agree to Agree: Neural Networks Share Classification Order on Real Datasets
Guy Hacohen, Leshem Choshen, Daphna Weinshall
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train independently initialized networks of the same architecture on the same real dataset | Neural networks on real classification benchmarks · ImageNet, CIFAR, MNIST, Fashion-MNIST, and Stack Overflow | similarity of example-wise learning order | Improved |
| Use a stronger architecture that can continue learning after matching examples learned by the weaker model | Neural networks with different architectures · ImageNet | similarity of learning order: 0.87 | Improved |
| Train on separate random partitions sampled from the same data distribution | Neural networks of the same architecture · Shared test set | similarity of accessibility-score order | Improved |
| Replace nonlinear layers with linear equivalents (ablation) | st-VGG · Small mammals | average accuracy: 0.56 -> 0.43 | Hurt |
| Replace nonlinear layers with linear equivalents (ablation) | st-VGG · Small mammals | shared learning-order behavior | No clear effect |
| Train neural networks on synthetic Gabor-patch data instead of a real benchmark dataset (ablation) | st-VGG · Gabor dataset | shared learning-order behavior | Hurt |
| Train on overlapping Gaussian classes (ablation) | Fully connected neural networks · Gaussian classification task | shared learning-order behavior | Hurt |
| Shuffle labels so that examples cannot generalize (ablation) | st-VGG · Small mammals with shuffled labels | shared learning-order behavior: 1.07 train; 1.35 test | Hurt |
| Use AdaBoost with linear weak classifiers instead of neural networks | CIFAR-10 classifiers · CIFAR-10 | learning-order correlation: 0.35 | Hurt |
| Give AdaBoost Inception-V3 penultimate-layer features (ablation) | AdaBoost compared with st-VGG · CIFAR-10 | learning-order correlation: 0.35 -> 0.05 | Hurt |
| Vary optimization, initialization, batch size, dropout, architecture, or activation settings (ablation) | Neural networks on real datasets · Multiple real benchmarks | qualitative learning-order similarity | No clear effect |
Closest papers by meaning
- When Deep Classifiers Agree: Analyzing Correlations between Learning Order and Image Statistics (ECCV 2022 · similarity 0.90)
- Neural Networks Learn Statistics of Increasing Complexity (ICML 2024 · similarity 0.86)
- Understanding and Leveraging the Learning Phases of Neural Networks (AAAI 2024 · similarity 0.85)
- When Representations Align: Universality in Representation Learning Dynamics (ICML 2024 · similarity 0.84)
- Three Mechanisms of Feature Learning in a Linear Network (ICLR 2025 · similarity 0.84)
- A Decade's Battle on Dataset Bias: Are We There Yet? (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
