Identifying good directions to escape the NTK regime and efficiently learn low-degree plus sparse polynomials

Keyphrases

• Sparse Polynomials 100%
• Neural Tangent Kernel 100%
• Sample Complexity 22%
• Gradient Descent 22%
• Second Order Term 22%
• Two-layer Neural Network 22%
• Neural Network 11%
• Class Function 11%
• Parameter Space 11%
• Ability to Learn 11%
• Gradient Method 11%
• Taylor Expansion 11%
• Spectral Analysis 11%
• Deep Learning 11%
• Minimax Optimal 11%
• Spectral Characterization 11%
• Parameter Vector 11%
• Target Function 11%
• Regularizer 11%
• Loss of Regularity 11%
• Convergence Guarantee 11%
• Generalization Guarantees 11%
• First-order Taylor Expansion 11%
• Global Minimizer 11%
• Lazy Training 11%

Mathematics

• Polynomial 100%
• Low Degree 100%
• Neural Network 60%
• Minimax 20%
• Parameter Space 20%
• Taylor Expansion 20%
• Parameter Vector 20%
• Order Taylor Expansion 20%
• Deep Learning 20%
• Spectral Characterization 20%
• Global Minimizer 20%

Computer Science

• Order Taylor Expansion 50%