Research map / Architectures and efficiency / RNNs and normalization
Limitations of Lazy Training of Two-layers Neural Network
Behrooz Ghorbani, Song Mei, Theodor Misiakiewicz, Andrea Montanari
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 |
|---|---|---|---|
| Train all network weights rather than use the neural tangent linearization | Two-layer network with quadratic activation · Quadratic functions (Gaussian features), N/d in (0,1) | prediction risk: R_{{\sf NN}, N}(f_*) \le R_{{\sf NT},N}(f_*)-{ m GAP}(\rho) | Improved |
| Train all network weights to select principal eigendirections | Two-layer network with quadratic activation · Quadratic functions (Gaussian features), N>d | prediction risk: zero risk | Improved |
| Use random features with fixed first-layer weights | Two-layer network · Quadratic functions (Gaussian features) | prediction risk: remains generally bounded away from zero | Hurt |
| Use random features with weights not aligned to the target quadratic | Random features · Quadratic functions (Gaussian features), ρ→∞ | normalized prediction risk: 1-\frac{\langle {\boldsymbol \Gamma},{\boldsymbol B}\rangle^2}{\|{\boldsymbol \Gamma}\|_F^2\|{\boldsymbol B}\|_F^2} | Hurt |
| Increase the number of NT neurons to at least the feature dimension | Neural tangent model · Quadratic functions (Gaussian features) | prediction risk: vanishes asymptotically for ρ ≥ 1 | Improved |
| Train all network weights rather than use RF or NT | Two-layer network · Mixture of Gaussians | prediction risk: strictly smaller | Improved |
| Increase the number of neurons in the mixture-of-Gaussians setting | Neural tangent model and fully trained network · Mixture of Gaussians | prediction risk: Bayes error is not achieved even for very wide networks N/d\gg 1 | No clear effect |
Closest papers by meaning
- A Theory of Non-Linear Feature Learning with One Gradient Step in Two-Layer Neural Networks (ICML 2024 · similarity 0.84)
- How does Gradient Descent Learn Features --- A Local Analysis for Regularized Two-Layer Neural Networks (NeurIPS 2024 · similarity 0.84)
- Grokking as the transition from lazy to rich training dynamics (ICLR 2024 · similarity 0.84)
- Asymptotic Analysis of Two-Layer Neural Networks after One Gradient Step under Gaussian Mixtures Data with Structure (ICLR 2025 · similarity 0.83)
- On the Hardness of Training Deep Neural Networks Discretely (AAAI 2025 · similarity 0.83)
- A New Neural Kernel Regime: The Inductive Bias of Multi-Task Learning (NeurIPS 2024 · similarity 0.83)
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.
