Abstract

In this talk, I will introduce a conceptually novel structured prediction model, GPstruct, which is kernelised, non-parametric, and supporting Bayesian posterior inference. GPstruct can be instantiated for a wide range of structured objects such as linear chain, tree, grid, and other general graphs. As a first proof of concept, the model is benchmarked on segmentation, chunking, and named entity recognition of text processing tasks and gesture segmentation of video processing task involving a linear chain structure. One of practical issues of GPstruct is the memory demand which is quadratic in the number of latent variables and training runtime that scales cubically. This prevents GPstruct from being applied to problems involving grid factor graphs, which are prevalent in computer vision applications. In the second part of the talk, I will describe a scaling trick based on ensemble learning, with weak learners (predictors) trained on subsets of the latent variables and bootstrap data, which can easily be distributed. We show experiments with 2 millions latent variables on image segmentation. Our method outperforms widely-used conditional random field models trained with pseudo-likelihood. Moreover, it improves over recent state-of-the-art marginal optimization methods in terms of predictive performance and uncertainty calibration. Finally, it generalizes well on all training set sizes.

Joint work with Sebastien Bratieres, Zoubin Ghahramani, and Sebastian Nowozin.