Our paper "Non-Stationary Spectral Kernels" is going to presented as a poster at NIPS 2017. The paper extends spectral mixture kernels into non-stationary kernels. Important features over previous proposals for non-stationary kernels include the ability to approximate any non-stationary kernel, and especially non-monotonic kernels, with some type of periodicity.
Another paper titled "A Mutually-Dependent Hadamard Kernel for Modelling Latent Variable Couplings" will also be presented at ACML 2017 both as oral as well as a poster presentation. This paper introduces a multi-output kernel that allows the outputs to have a non-stationary, or input-dependent, correlation structure, in contrast to popular Kronecker structured kernels. While this kernel could be used in plain multi-output prediction, we also deploy the kernel in a latent variable model.
These two papers also constitute the two most important papers to be included in my doctoral dissertation, which I hope will be finished by the end of the year!