Core Concepts
A novel learning strategy that models the sought function as a sample function of a non-stationary Gaussian Process, which nests multiple stationary Gaussian Processes within it, to effectively capture inhomogeneities in the correlation structure of the available data.
Abstract
The content presents a new learning strategy for modeling the functional relationship between a pair of variables, while addressing the challenge of inhomogeneities in the correlation structure of the available data. The key aspects of the proposed approach are:
The sought function is modeled as a sample function of a non-stationary Gaussian Process (GP), which nests within itself multiple other GPs.
It is proven that the nested GPs can be stationary, thereby establishing the sufficiency of two GP layers.
The non-stationary kernel is designed such that each hyperparameter is dependent on the sample function drawn from the outer non-stationary GP, allowing the hyperparameters to vary with the input locations.
To make the model computationally feasible, the authors show that the average effect of drawing different sample functions from the non-stationary GP is equivalent to drawing a sample function from each of a set of stationary GPs, with the hyperparameters updated during the inference process.
The kernel is fully non-parametric, requiring the learning of only one hyperparameter per layer of GP, for each dimension of the input variable.
The proposed learning strategy is illustrated on a real-world dataset, and its predictive performance is compared against various existing non-stationary and stationary kernel models, as well as Deep Neural Networks.
Stats
The dataset used in the empirical illustration is a real-world dataset, but the specific details about the variables and their units are not provided in the content.
Quotes
No striking quotes were identified in the content.