Automated Augmented Conjugate Inference for Non-conjugate Gaussian Process Models

We propose automated augmented conjugate inference, a new inference method\nfor non-conjugate Gaussian processes (GP) models. Our method automatically\nconstructs an auxiliary variable augmentation that renders the GP model\nconditionally conjugate. Building on the conjugate structure of the augmented\nmodel, we develop two inference methods. First, a fast and scalable stochastic\nvariational inference method that uses efficient block coordinate ascent\nupdates, which are computed in closed form. Second, an asymptotically correct\nGibbs sampler that is useful for small datasets. Our experiments show that our\nmethod are up two orders of magnitude faster and more robust than existing\nstate-of-the-art black-box methods.\n