I think it still is. The output is a function fitting the data, so as long as that function is amenable to analysis, it doesn't really matter how it's found. I

I think it still is. The output is a function fitting the data, so as long as that function is amenable to analysis, it doesn't really matter how it's found. In fact, the paper discusses some modifications they made (reducing the number of parameters) to make the output kernels more intelligible to humans.