Research map / Optimization / Implicit bias and generalization
Visualizing the Loss Landscape of Neural Nets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 filter-wise normalized directions instead of unnormalized directions (ablation) | loss-landscape visualization · CIFAR-10 classifiers | correlation of sharpness with test error | Helped |
| use filter-wise normalization instead of layer-wise normalization (ablation) | loss-landscape visualization · VGG-9, CIFAR-10 | correlation of sharpness with test error | Helped |
| use unnormalized directions instead of filter-wise normalization (ablation) | loss-landscape visualization · VGG-9, CIFAR-10 | correlation of sharpness with test error | Hurt |
| add shortcut connections (ablation) | ResNet-56 · CIFAR-10 | test error: 13.31 -> 5.89 | Helped |
| add shortcut connections (ablation) | ResNet-110 · CIFAR-10 | test error: 16.44 -> 5.79 | Helped |
| add shortcut connections (ablation) | ResNet-20 · CIFAR-10 | test error: 8.18 -> 7.37 | Helped |
| increase batch size from 128 to 4096 (ablation) | ResNet-56 with SGD, WD = 0 · CIFAR-10 | test error: 8.26 -> 13.93 | Hurt |
| increase batch size from 128 to 4096 (ablation) | ResNet-56 with Adam, WD = 0 · CIFAR-10 | test error: 9.55 -> 14.30 | Hurt |
| increase batch size from 128 to 4096 (ablation) | ResNet-56 with SGD, WD = 5e-4 · CIFAR-10 | test error: 5.89 -> 10.59 | Hurt |
| increase batch size from 128 to 4096 (ablation) | ResNet-56 with Adam, WD = 5e-4 · CIFAR-10 | test error: 7.67 -> 12.36 | Hurt |
| add weight decay 5e-4 (ablation) | ResNet-56 with SGD, bs=128 · CIFAR-10 | test error: 8.26 -> 5.89 | Helped |
| add weight decay 5e-4 (ablation) | ResNet-56 with SGD, bs=4096 · CIFAR-10 | test error: 13.93 -> 10.59 | Helped |
Closest papers by meaning
- Large Scale Structure of Neural Network Loss Landscapes (NeurIPS 2019 · similarity 0.91)
- Exploring The Loss Landscape Of Regularized Neural Networks Via Convex Duality (ICLR 2025 · similarity 0.89)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.88)
- Neuro-Visualizer: A Novel Auto-Encoder-Based Loss Landscape Visualization Method With an Application in Knowledge-Guided Machine Learning (ICML 2024 · similarity 0.88)
- Flat Channels to Infinity in Neural Loss Landscapes (NeurIPS 2025 · similarity 0.88)
- Deeper or Wider: A Perspective from Optimal Generalization Error with Sobolev Loss (ICML 2024 · similarity 0.87)
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.