Bernstein-Schur Kernels: Random Features by Sketched Modulation and Radial Randomization
Researchers have introduced Bernstein-Schur kernels, a novel approach to creating nonstationary kernels by combining finite-feature kernels with completely monotone shift-invariant kernels. This method utilizes a dual randomization strategy, sketching the modulation factor and randomizing the radial factor. The proposed technique offers an explicit finite-dimensional feature map construction, reducing the feature dimension compared to existing methods and providing theoretical guarantees on unbiasedness and operator-norm bounds. AI
IMPACT Introduces a new kernel construction that could enhance machine learning model performance on complex datasets.