Research map / Optimization / Implicit bias and generalization
Optimizer Choice Matters For The Emergence of Neural Collapse
Jim Zhao, Tin Sum Cheng, Wojciech Masarczyk, Aurelien Lucchi
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 coupled weight decay (Adam) instead of decoupled (AdamW) (ablation) | Adam family · Final NC metrics, main setup | NC0: 5.33 -> 0.34 | Improved |
| use coupled weight decay (Adam) instead of decoupled (AdamW) (ablation) | Adam family · Final NC metrics, main setup | NC3: 0.78 -> 0.29 | Improved |
| use coupled weight decay (Signum) instead of decoupled (SignumW) (ablation) | sign-based optimizers · Final NC metrics, main setup | NC0: 3185.69 -> 0.78 | Improved |
| use coupled weight decay (SGD) instead of decoupled (SGDW) (ablation) | SGD · Final NC metrics, main setup | NC0: 0.55 -> 2.14e-04 | Improved |
| increase coupled share of weight decay while total fixed at 0.0005 (AdamW to Adam) (ablation) | Adam/AdamW interpolation · ResNet9 on MNIST | NC0, NC2, NC3 | Improved |
| interpolate Adam/AdamW weight decay coupling (ablation) | Adam/AdamW interpolation · ResNet9 on MNIST | validation accuracy | No clear effect |
| add weight decay (ablation) | all optimizers · ResNet9/VGG9 on MNIST, FashionMNIST, CIFAR10 | NC metrics | Improved |
| increase momentum at fixed nonzero weight decay (ablation) | SGD · ResNet9 on FashionMNIST | NC0, NC2, NC3 | Improved |
| increase momentum (0.7 vs 0.9) with matching train loss (ablation) | SGD · ResNet9 on FashionMNIST | NC metrics at equal train loss | Improved |
| train 200 vs 2000 epochs (ablation) | Adam/AdamW · ResNet9 on FashionMNIST | NC metrics | No clear effect |
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)
- Wide Neural Networks Trained with Weight Decay Provably Exhibit Neural Collapse (ICLR 2025 · similarity 0.89)
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.88)
- Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks (ICML 2025 · similarity 0.88)
- Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling (NeurIPS 2024 · similarity 0.87)
- Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking (NeurIPS 2025 · 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.
