← main page
3. Kernel Regression
03a · Rotation-invariant kernels
03b · KRR on CIFAR-10
\[ \Theta(\mathbf{x}, \mathbf{x}') = \Theta\!\left(\frac{\|\mathbf{x}\|}{\sqrt d},\ \frac{\|\mathbf{x}'\|}{\sqrt d},\ \frac{\mathbf{x}^\top \mathbf{x}'}{d}\right) \]
\(\theta\)
\(\Theta(1, 1, \cos\theta)\)
Gaussian
Laplacian
ReLU NNGP
ReLU NTK
tanh NNGP
tanh NTK
linear
domain \(\mathcal{X} = S^1\)
\(\theta\)
\(\phi_i(\theta)\)
\(i\)
\(\lambda_i\)
training set size \(n\)
test MSE
Gaussian
Laplacian
ReLU NNGP
ReLU NTK
tanh NNGP
tanh NTK
linear
go
show theory