Research map / Optimization / Implicit bias and generalization
Early Stage Convergence and Global Convergence of Training Mildly Parameterized Neural Networks
Mingze Wang, Chao Ma
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 |
|---|---|---|---|
| increase network width m from 100 to 1000 (eta=0.01) | two-layer ReLU network, GD · CIFAR-10 first 1000 | realistic loss descent: 3.45\times10^{-1} -> 6.53\times10^{-1} | Improved |
| increase network width m from 100 to 1000 (eta=0.01) | two-layer ReLU network, GD · MNIST first 1000 | realistic loss descent: 3.89\times10^{-1} -> 6.45\times10^{-1} | Improved |
| reduce learning rate from 0.01 to 0.001 (m=200) | two-layer ReLU network, GD · CIFAR-10 first 1000 | realistic loss descent: 4.89\times10^{-1} -> 5.37\times10^{-1} | No clear effect |
| reduce learning rate from 0.01 to 0.001 (m=200) | two-layer ReLU network, GD · MNIST first 1000 | realistic loss descent: 4.78\times10^{-1} -> 4.92\times10^{-1} | No clear effect |
| reduce learning rate from 0.01 to 0.001 | early-stage bound · hitting iteration | iterations to reach bound: 34 -> 346 | Hurt |
| change dataset between MNIST and CIFAR-10 | theoretical loss descent bound · loss descent bound | bound value: 2.63\times10^{-1} | No clear effect |
Closest papers by meaning
- An Improved Analysis of Training Over-parameterized Deep Neural Networks (NeurIPS 2019 · similarity 0.90)
- A proof of convergence for the gradient descent optimization method with random initializations in the training of neural networks with ReLU activation for piecewise linear target functions (JMLR 2022 · similarity 0.90)
- Fast Convergence of Natural Gradient Descent for Over-Parameterized Neural Networks (NeurIPS 2019 · similarity 0.90)
- Fast Convergence of Natural Gradient Descent for Over-parameterized Physics-Informed Neural Networks (ICLR 2026 · similarity 0.89)
- Convergence Analysis of Newton's Method for Neural Networks in the Overparameterized Limit (NeurIPS 2026 · similarity 0.89)
- On the Unstable Convergence Regime of Gradient Descent (AAAI 2024 · similarity 0.89)
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.
