Research map / Optimization / Implicit bias and generalization
On the Proof of Global Convergence of Gradient Descent for Deep ReLU Networks with Linear Widths
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 the output-layer gradient and track the last hidden-layer output instead of tracking activation-pattern changes across layers | Gradient descent for deep ReLU networks · Training set (general data) | last hidden layer width sufficient for global convergence: \Omega(N^8) -> n_{L-1}≥N | Helped |
| Apply the sufficient initialization conditions under LeCun initialization for a two-layer network | Two-layer ReLU network · Training set (LeCun initialization) | last hidden layer width sufficient for global convergence: n_1=\Omega(N^2) | Helped |
| Use the specified smaller output-weight variance for deep networks with LeCun hidden-weight initialization | Deep ReLU network · Training set (stated data-distribution assumptions) | last hidden layer width sufficient for global convergence: n_{L-1}=\bigTildeOmg{N^3\mathop{\rm max}\nolimits(m,n_0)^3} | Helped |
| Train all layers with standard gradient descent rather than fixing the output layer and optimizing only hidden layers | Deep ReLU network · Training set (general data) | global convergence guarantee | Helped |
| Use a single wide last hidden layer while allowing the remaining layers to have constant widths | Deep ReLU network · Training set (initializations satisfying Theorem 1) | hidden-layer width requirement: n_{L-1}≥N | Helped |
Closest papers by meaning
- Convergence of Near-Linear Width ReLU Networks with Unbalanced Initialization (NeurIPS 2026 · similarity 0.92)
- Convergence of the Gradient Flow for Shallow ReLU Networks on Weakly Interacting Data (NeurIPS 2025 · similarity 0.92)
- Gradient Descent Provably Optimizes Over-parameterized Neural Networks (ICLR 2019 · similarity 0.92)
- Convex Relaxations of ReLU Neural Networks Approximate Global Optima in Polynomial Time (ICML 2024 · similarity 0.90)
- Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training (ICML 2025 · similarity 0.90)
- Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks (NeurIPS 2025 · similarity 0.90)
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.