Research map / Optimization / Implicit bias and generalization
Direction Matters: On the Implicit Bias of Stochastic Gradient Descent with Moderate Learning Rate
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 moderate initial learning rate followed by annealing instead of GD | Overparameterized linear regression · Linear regression, moderate learning rate | convergence direction | Helped |
| Use moderate initial learning rate followed by annealing instead of GD | Overparameterized linear regression with early stopping · Linear regression, early-stopped | estimation error | Helped |
| Use moderate learning rate instead of small learning rate (ablation) | SGD · Overparameterized linear regression, early-stopped | estimation error | Helped |
| Use moderate learning rate instead of small learning rate (ablation) | SGD · FashionMNIST subset | test accuracy | Helped |
| Use small learning rate instead of moderate learning rate (ablation) | SGD · FashionMNIST subset | test accuracy | Hurt |
| Use moderate learning rate instead of small learning rate (ablation) | GD · FashionMNIST subset | test accuracy | No clear effect |
| Scale the initial learning rate linearly with batch size | SGD · Overparameterized linear regression | directional bias | Helped |
| Continue SGD with a moderate learning rate even when training loss appears stalled | SGD · Overparameterized linear regression | convergence direction | Helped |
Closest papers by meaning
- A Guide Through the Zoo of Biased SGD (NeurIPS 2023 · similarity 0.92)
- Risk Bounds of Accelerated SGD for Overparameterized Linear Regression (ICLR 2024 · similarity 0.90)
- Bias of Stochastic Gradient Descent or the Architecture: Disentangling the Effects of Overparameterization of Neural Networks (ICML 2024 · similarity 0.90)
- Non-asymptotic Analysis of Biased Adaptive Stochastic Approximation (NeurIPS 2024 · similarity 0.90)
- Beyond Implicit Bias: The Insignificance of SGD Noise in Online Learning (ICML 2024 · similarity 0.90)
- Tighter Convergence Bounds for Shuffled SGD via Primal-Dual Perspective (NeurIPS 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.