Research map / Learning theory and algorithms / Kernels and regression theory
Relative stability toward diffeomorphisms indicates performance in deep nets
Leonardo Petrini, Alessandro Favero, Mario Geiger, Matthieu Wyart
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 |
|---|---|---|---|
| use R_f instead of D_f alone to predict test error (ablation) | trained nets · FashionMNIST | correlation with test error: 0.72 -> 0.94 | Improved |
| use R_f instead of D_f alone (ablation) | trained nets · MNIST | correlation with test error: 0.71 -> 0.75 | Improved |
| use R_f instead of D_f alone (ablation) | trained nets · SVHN | correlation with test error: 0.87 -> 0.81 | No clear effect |
| use R_f instead of D_f alone (ablation) | trained nets · Tiny ImageNet | correlation with test error: 0.69 -> 0.74 | Improved |
| use R_f instead of G_f alone (ablation) | trained nets · FashionMNIST | correlation with test error: -0.68 -> 0.94 | Improved |
| train EfficientNetB0 by transfer learning from ImageNet instead of from scratch | EfficientNetB0 · CIFAR10 | test error and R_f | Improved |
| increase training set size P | trained nets · MNIST/FashionMNIST/CIFAR10 | stability to additive noise G_f | Hurt |
| increase training set size P | ResNet/EfficientNet · MNIST/FashionMNIST/CIFAR10 | stability to diffeomorphisms D_f | Mixed |
| use primitive CNN LeNet instead of SOTA nets | LeNet · CIFAR10 | R_f: 0.6 | Hurt |
| change interpolation method for image deformations (ablation) | diffeomorphism probes · MNIST/FashionMNIST | R_f | No clear effect |
Closest papers by meaning
- Invariance and Stability of Deep Convolutional Representations (NeurIPS 2017 · similarity 0.87)
- A Unified Stability Analysis of SAM vs SGD: Role of Data Coherence and Emergence of Simplicity Bias (NeurIPS 2025 · similarity 0.86)
- On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling (NeurIPS 2025 · similarity 0.86)
- DySLIM: Dynamics Stable Learning by Invariant Measure for Chaotic Systems (ICML 2024 · similarity 0.85)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.85)
- Stable Minima of ReLU Neural Networks Suffer from the Curse of Dimensionality: The Neural Shattering Phenomenon (NeurIPS 2025 · similarity 0.85)
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.
