Transformed Gaussian Process

Transformed Gaussian Processes (TGPs) enhance the flexibility of standard Gaussian Processes by applying invertible transformations to their samples, enabling modeling of more complex, non-stationary data. Current research focuses on developing scalable inference methods for TGPs, including variational inference techniques and efficient sparse approximations, and extending the architecture to deep (DTGPs) and multi-class (ETGPs) settings. These advancements improve the accuracy and efficiency of TGPs for regression and classification tasks, particularly in high-dimensional or multi-class problems, offering a powerful alternative to traditional Gaussian Process models.