4.3 Neural Tangent Kernel (NTK) — Spectral Decomposition of Gradient Flow

We turn gradient descent into a continuous-time model, then use the NTK spectrum to explain why some parts of a function are learned faster than others.

Watch the full series:
https://www.youtube.com/playlist?list=PLdMM2W7mxqIYDURvh37jLz8sd8sDjonk5

0:00 — NTK dynamics and ODE
6:03 — Spectral decomposition and training dynamics
14:28 — Interpreting the spectrum

Topics:
• Gradient flow (ODE)
• Spectral / eigen decomposition
• Learning dynamics
• Fast vs slow modes
• PCA / covariance view
• Kernel perspective (NTK)
• Limits of fixed features
• SGD as noisy kernel

Видео 4.3 Neural Tangent Kernel (NTK) — Spectral Decomposition of Gradient Flow канала Jingyuan Hu