Gaussian processes (GPs) are powerful but struggle with high-dimensional data like images. A standard approach, the fully anisotropic radial basis function (RBF) kernel, assigns a separate hyperparameter to each input dimension, which quickly becomes impractical as image resolution grows. A new paper on arXiv proposes a structured alternative.
The authors introduce RBF kernels that use structured hyperparameter sharing for image inputs. Instead of treating every pixel independently, the kernel shares parameters in a way that reflects the structure of image data, and the title suggests a channel-additive design. This reduces the number of free hyperparameters while preserving flexibility.
The work aims to improve relevance determination—the ability of a GP to identify which input features matter most. By structuring the kernel, the model can better handle images without the computational and statistical costs of fully anisotropic kernels. The abstract does not include experimental results, so the practical gains are not yet quantified, but the approach offers a promising direction for GP-based image analysis.