Research map / Learning theory and algorithms / Matrix and tensor methods
Blessing of Depth in Linear Regression: Deeper Models Have Flatter Landscape Around the True Solution
Jianhao Ma, Salar Fattahi
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 |
|---|---|---|---|
| replace cross-entropy with l1-loss | ResNet-18 · CIFAR-10 10% label noise | test accuracy: 91.52% -> 94.16% | Improved |
| replace cross-entropy with l1-loss | ResNet-50 · CIFAR-10 10% label noise | test accuracy: 90.87% -> 92.68% | Improved |
| replace cross-entropy with l1-loss | ResNet-18 · CIFAR-100 10% label noise | test accuracy: 70.17% -> 73.14% | Improved |
| replace cross-entropy with l1-loss | ResNet-50 · CIFAR-100 10% label noise | test accuracy: 71.30% -> 76.46% | Improved |
| increase depth from ResNet-18 to ResNet-50 with l1-loss | ResNet l1-loss · CIFAR-100 10% label noise | test accuracy: 73.14% -> 76.46% | Improved |
| increase depth from ResNet-18 to ResNet-50 with l1-loss | ResNet l1-loss · CIFAR-10 10% label noise | test accuracy: 94.16% -> 92.68% | Hurt |
| increase depth from ResNet-18 to ResNet-50 with CE loss | ResNet CE · CIFAR-10 10% label noise | test accuracy: 91.52% -> 90.87% | No clear effect |
| use 4-layer instead of 2- or 3-layer model | SubGM on deep matrix recovery · synthetic, 5% corrupted | generalization error | Improved |
| use 3-layer instead of 2-layer model | SubGM on diagonal linear regression · synthetic robust regression | generalization error | Improved |
| increase depth N | SubGM on diagonal linear regression · synthetic robust regression | convergence speed | Hurt |
| use 4-layer ReLU model with l1-loss | SubGM · synthetic sin(theta^T x), 5% corrupted | model-mismatch error | Improved |
Closest papers by meaning
- Deep linear networks for regression are implicitly regularized towards flat minima (NeurIPS 2024 · similarity 0.89)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · similarity 0.87)
- Implicit Sparse Regularization: The Impact of Depth and Early Stopping (NeurIPS 2021 · similarity 0.87)
- Large Scale Structure of Neural Network Loss Landscapes (NeurIPS 2019 · similarity 0.87)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.85)
- Private Gradient Descent for Linear Regression: Tighter Error Bounds and Instance-Specific Uncertainty Estimation (ICML 2024 · 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.
