Efficient Multioutput Gaussian Processes through Variational Inducing Kernels.

Journal of Machine Learning Research - Proceedings Track 01/2010; 9:25-32.
Source: DBLP


Interest in multioutput kernel methods is increas- ing, whether under the guise of multitask learn- ing, multisensor networks or structured output data. From the Gaussian process perspective a multioutput Mercer kernel is a covariance func- tion over correlated output functions. One way of constructing such kernels is based on convolution processes (CP). A key problem for this approach is efficient inference. ´ Alvarez and Lawrence re- cently presented a sparse approximation for CPs that enabled efficient inference. In this paper, we extend this work in two directions: we in- troduce the concept of variational inducing func- tions to handle potential non-smooth functions involved in the kernel CP construction and we consider an alternative approach to approximate inference based on variational methods, extend- ing the work by Titsias (2009) to the multiple output case. We demonstrate our approaches on prediction of school marks, compiler perfor- mance and financial time series.

Download full-text


Available from: David Luengo
  • Source
    • "In contrast to most other multi-output models, the aim of a MTGP is to learn the correlation between and within the tasks concurrently, and thereby to improve overall modelling accuracy compared to a STGP. MTGPs have been investigated for applications such as the analysis of compiler performance [10], robotic inverse dynamics [11], financial time-series [12], and environmental sensor networks [13]. We note in passing that a multi-task GP is sometimes called a " multi-output " GP in the literature. "
    [Show abstract] [Hide abstract]
    ABSTRACT: In extracranial robotic radiotherapy, tumour motion is compensated by tracking external and internal surrogates. To compensate system specific time delays, time series prediction of the external optical surrogates is used. We investigate whether the prediction accuracy can be increased by expanding the current clinical setup by an accelerometer, a strain belt and a flow sensor. Four previously published prediction algorithms are adapted to multivariate inputs-normalized least mean squares (nLMS), wavelet-based least mean squares (wLMS), support vector regression (SVR) and relevance vector machines (RVM)-and evaluated for three different prediction horizons. The measurement involves 18 subjects and consists of two phases, focusing on long term trends (M1) and breathing artefacts (M2). To select the most relevant and least redundant sensors, a sequential forward selection (SFS) method is proposed. Using a multivariate setting, the results show that the clinically used nLMS algorithm is susceptible to large outliers. In the case of irregular breathing (M2), the mean root mean square error (RMSE) of a univariate nLMS algorithm is 0.66 mm and can be decreased to 0.46 mm by a multivariate RVM model (best algorithm on average). To investigate the full potential of this approach, the optimal sensor combination was also estimated on the complete test set. The results indicate that a further decrease in RMSE is possible for RVM (to 0.42 mm). This motivates further research about sensor selection methods. Besides the optical surrogates, the sensors most frequently selected by the algorithms are the accelerometer and the strain belt. These sensors could be easily integrated in the current clinical setup and would allow a more precise motion compensation.
    Full-text · Article · Sep 2014 · Physics in Medicine and Biology
  • Source
    • "1.2, 2.11).Finally, each output dimension of the model is treated with an independent GP, which many not make good use of the data. Although methods for efficient dependent GPs exist [3], they have not yet been applied to reinforcement learning. To make decisions using our model, we employ Thompson sampling [38] [39]. "
    [Show abstract] [Hide abstract]
    ABSTRACT: This paper proposes an online tree-based Bayesian approach for reinforcement learning. For inference, we employ a generalised context tree model. This defines a distribution on multivariate Gaussian piecewise-linear models, which can be updated in closed form. The tree structure itself is constructed using the cover tree method, which remains efficient in high dimensional spaces. We combine the model with Thompson sampling and approximate dynamic programming to obtain effective exploration policies in unknown environments. The flexibility and computational simplicity of the model render it suitable for many reinforcement learning problems in continuous state spaces. We demonstrate this in an experimental comparison with least squares policy iteration.
    Full-text · Article · May 2013 · Journal of Machine Learning Research
  • Source
    • "In this approach, the sample points X m are chosen to maximize a variational lower bound on the marginal likelihood, therefore providing a clear methodology for the choice of the support set. Following their idea, ´ Alvarez et al. (2010) proposed the variational inference for sparse convolved multiple output GPs. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Multi-task learning models using Gaussian processes (GP) have been recently developed and successfully applied in various applications. The main difficulty with this approach is the computational cost of inference using the union of examples from all tasks. The paper investigates this problem for the grouped mixed-effect GP model where each individual response is given by a fixed-effect, taken from one of a set of unknown groups, plus a random individual effect function that captures variations among individuals. Such models have been widely used in previous work but no sparse solutions have been developed. The paper presents the first sparse solution for such problems, showing how the sparse approximation can be obtained by maximizing a variational lower bound on the marginal likelihood, generalizing ideas from single-task Gaussian processes to handle the mixed-effect model as well as grouping. Experiments using artificial and real data validate the approach showing that it can recover the performance of inference with the full sample, that it outperforms baseline methods, and that it outperforms state of the art sparse solutions for other multi-task GP formulations.
    Full-text · Conference Paper · Sep 2012
Show more