diff --git a/assets/rbf_unit_sphere.svg b/assets/rbf_unit_sphere.svg new file mode 100644 index 0000000..51b9f2f --- /dev/null +++ b/assets/rbf_unit_sphere.svg @@ -0,0 +1 @@ +Normalize input gradients to unit vectors.A scalar adaptively adjusted to prevent kernel vanishing.Diagonal matrix containing the lengthscales. \ No newline at end of file diff --git a/index.md b/index.md index e2ab614..d985306 100644 --- a/index.md +++ b/index.md @@ -45,4 +45,8 @@ Please cite our work if you find it useful: Given a base kernel \\(\kappa\\), we define the kernel in the input-gradient space for a minibatch of training samples \\(\mathcal{B}=\\{(\mathbf{x}\_b, y\_b\\}\_{b=1}^B\\) as follows: -drawing \ No newline at end of file +drawing + +We choose the RBF kernel on a unit sphere as the base kernel \\(\kappa\\): + +drawing \ No newline at end of file