Research map / Architectures and efficiency / RNNs and normalization
Taxonomizing local versus global structure in neural network loss landscapes
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 |
|---|---|---|---|
| Reach a globally well-connected, locally flat landscape with high CKA similarity | ResNet18 · CIFAR-10 | test accuracy | Helped |
| Increase model width | ResNet18 · CIFAR-10 | mode connectivity | Helped |
| Increase the amount of training data | ResNet18 · CIFAR-10 | CKA similarity | Helped |
| Increase training label noise | ResNet18 · CIFAR-10 | mode connectivity | Hurt |
| Train to approximately zero loss | ResNet18 in a globally poorly-connected region · CIFAR-10 with 10% randomized labels | test accuracy | Hurt |
| Add learning rate decay during training (ablation) | ResNet18 · CIFAR-10 | test accuracy | Helped |
| Use weight decay as the temperature-like parameter instead of batch size (ablation) | ResNet18 · CIFAR-10 | test accuracy | No clear effect |
| Increase the amount of training data (ablation) | ResNet18 · CIFAR-10 | test accuracy | Helped |
| Increase the amount of training data (ablation) | ResNet18 · CIFAR-10 | Hessian trace | Mixed |
| Increase the proportion of randomized training labels (ablation) | ResNet18 · CIFAR-10 | Hessian trace | No clear effect |
| Increase batch size (ablation) | ResNet18 · CIFAR-10 | local sharpness | Hurt |
| Use a small model with poor connectivity and train to zero loss (ablation) | Transformer · IWSLT 2016 De-En | generalization | Hurt |
Closest papers by meaning
- Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime (ICML 2025 · similarity 0.89)
- Super Consistency of Neural Network Landscapes and Learning Rate Transfer (NeurIPS 2024 · similarity 0.89)
- A Dynamical Model of Neural Scaling Laws (ICML 2024 · similarity 0.87)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · similarity 0.87)
- Flat Channels to Infinity in Neural Loss Landscapes (NeurIPS 2025 · similarity 0.87)
- Neuro-Visualizer: A Novel Auto-Encoder-Based Loss Landscape Visualization Method With an Application in Knowledge-Guided Machine Learning (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.