Hierarchical mixtures of experts and the EM algorithm
Neural Computation
Predictive Approaches for Choosing Hyperparameters in Gaussian Processes
Neural Computation
Gaussian Processes for Machine Learning (Adaptive Computation and Machine Learning)
Gaussian Processes for Machine Learning (Adaptive Computation and Machine Learning)
Curve prediction and clustering with mixtures of Gaussian process functional regression models
Statistics and Computing
Hi-index | 0.00 |
The mixture of Gaussian processes (MGP) is an important probabilistic model which is often applied to the regression and classification of temporal data. But the existing EM algorithms for its parameter learning encounters a hard difficulty on how to compute the expectations of those assignment variables (as the hidden ones). In this paper, we utilize the leave-one-out cross-validation probability decomposition for the conditional probability and develop an efficient EM algorithm for the MGP model in which the expectations of the assignment variables can be solved directly in the E-step. In the M-step, a conjugate gradient method under a standard Wolfe-Powell line search is implemented to learn the parameters. Furthermore, the proposed EM algorithm can be carried out in a hard cutting way such that each data point is assigned to the GP expert with the highest posterior in the E-step and then the parameters of each GP expert can be learned with these assigned data points in the M-step. Therefore, it has a potential advantage of handling large datasets in comparison with those soft cutting methods. The experimental results demonstrate that our proposed EM algorithm is effective and efficient.