Research map / Learning theory and algorithms / Prediction and decision losses
The Goldilocks Zone: Towards Better Understanding of Neural Network Loss Landscapes
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 |
|---|---|---|---|
| Place the random hypersurface within the Goldilocks zone | Optimization on random low-dimensional hypersurfaces · MNIST and CIFAR-10 | accuracy | Helped |
| Move the initialization radius above the Xavier radius | Optimization on random low-dimensional hyperplanes · MNIST and CIFAR-10 | accuracy | Hurt |
| Move the initialization radius away from the Xavier radius | Optimization on random low-dimensional hyperspheres · MNIST and CIFAR-10 | accuracy | Hurt |
| Increase the dimension of the random hyperplane | Random low-dimensional hyperplanes · MNIST and CIFAR-10 | fraction of positive Hessian eigenvalues | Hurt |
| Increase the dimension of the random hyperplane | Random low-dimensional hyperplanes · MNIST and CIFAR-10 | Tr(H)/||H|| | No clear effect |
| Select initializations with a higher fraction of positive Hessian eigenvalues | Full-space training from Xavier-initialized points · MNIST | training speed | Helped |
| Select initializations with a higher Hessian trace | Full-space training from Xavier-initialized points · MNIST | training speed | Helped |
| Select the initialization with the lowest initial validation loss | Full-space training from Xavier-initialized points · MNIST | training speed | Helped |
| Use tanh instead of ReLU | Fully connected networks · MNIST and CIFAR-10 | strength of Goldilocks-zone curvature effect | Hurt |
| Use convolutional networks instead of fully connected networks | Neural network loss landscapes · MNIST and CIFAR-10 | Tr(H)/||H|| | Mixed |
Closest papers by meaning
- Deconstructing the Goldilocks Zone of Neural Network Initialization (ICML 2024 · similarity 0.94)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.88)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · similarity 0.88)
- Super Consistency of Neural Network Landscapes and Learning Rate Transfer (NeurIPS 2024 · similarity 0.88)
- Flat Channels to Infinity in Neural Loss Landscapes (NeurIPS 2025 · similarity 0.87)
- Simplicity Bias via Global Convergence of Sharpness Minimization (ICML 2024 · 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.