These are the centers of your radial basis functions. Each circle corresponds to a Gaussian G(x;m,s)
with mean m
and (scalar) variance s
. Here the mean varies from x=-5
to x=5
whereas |y|=2
is held fixed. The variance seems to be isotropic and identical across all Gaussians.
Using these functions, one can then go on and do some kind of regression. There are multiple alternatives for this: one can for example consider the simple linear superposition of those Gaussians and optimize the superposition coefficients (--that would be linear regression-style). On the other hand, one can also try to optimize the hyper-parameters, namely the means and variances, so that the fit becomes closest.
The basic problem, however, already becomes clear from this figure: the higher the dimension, the more functions you need in order to obtain a good fit (for example, think about interpolating something which is located around the point (-5,-5)
-- it simply doesn't work). This problem is often termed the curse of dimensionality.