NTK Kernel Regression Demo

Click to add training points. The infinite-width ReLU NTK is computed in closed form, and the predictive mean is the kernel ridge regression solution.

NTK posterior mean ±1σ band training points

Depth L 3

Bias β 0.10

Noise σ² 0.010

Show NN samples

Width n 200

The blue curve is the predictive mean of kernel ridge regression with the limiting ReLU NTK at depth 3. The grey band is one standard deviation under the NTK Gaussian posterior. Toggle "Show NN samples" to overlay finite-width networks drawn at the chosen width.

Kernel matrix on test grid

Heatmap of Θ_∞^(L)(x, x') across the input grid. Brighter means more correlated.

Neural Tangent Kernel: Regression Demo