In this doctoral thesis, I present my research into applying machine learning techniques for reconstructing species interaction networks in ecology, reconstructing molecular signalling pathways and gene regulatory networks in systems biology, and inferring parameters in ordinary differential equation (ODE) models of signalling pathways. Together, the methods I have developed for these applications demonstrate the usefulness of machine learning for reconstructing networks and inferring network parameters from data. The thesis consists of three parts. The first part is a detailed comparison of applying static Bayesian networks, relevance vector machines, and linear regression with L1 regularisation (LASSO) to the problem of reconstructing species interaction networks from species absence/presence data in ecology (Faisal et al., 2010). I describe how I generated data from a stochastic population model to test the different methods and how the simulation study led us to introduce spatial autocorrelation as an important covariate. I also show how we used the results of the simulation study to apply the methods to presence/absence data of bird species from the European Bird Atlas. The second part of the thesis describes a time-varying, non-homogeneous dynamic Bayesian network model for reconstructing signalling pathways and gene regulatory networks, based on L`ebre et al. (2010). I show how my work has extended this model to incorporate different types of hierarchical Bayesian information sharing priors and different coupling strategies among nodes in the network. The introduction of these priors reduces the inference uncertainty by putting a penalty on the number of structure changes among network segments separated by inferred changepoints (Dondelinger et al., 2010; Husmeier et al., 2010; Dondelinger et al., 2012b). Using both synthetic and real data, I demonstrate that using information sharing priors leads to a better reconstruction accuracy of the underlying gene regulatory networks, and I compare the different priors and coupling strategies. I show the results of applying the model to gene expression datasets from Drosophila melanogaster and Arabidopsis thaliana, as well as to a synthetic biology gene expression dataset from Saccharomyces cerevisiae. In each case, the underlying network is time-varying; for Drosophila melanogaster, as a consequence of measuring gene expression during different developmental stages; for Arabidopsis thaliana, as a consequence of measuring gene expression for circadian clock genes under different conditions; and for the synthetic biology dataset, as a consequence of changing the growth environment. I show that in addition to inferring sensible network structures, the model also successfully predicts the locations of changepoints. The third and final part of this thesis is concerned with parameter inference in ODE models of biological systems. This problem is of interest to systems biology researchers, as kinetic reaction parameters can often not be measured, or can only be estimated imprecisely from experimental data. Due to the cost of numerically solving the ODE system after each parameter adaptation, this is a computationally challenging problem. Gradient matching techniques circumvent this problem by directly fitting the derivatives of the ODE to the slope of an interpolant. I present an inference procedure for a model using nonparametric Bayesian statistics with Gaussian processes, based on Calderhead et al. (2008). I show that the new inference procedure improves on the original formulation in Calderhead et al. (2008) and I present the result of applying it to ODE models of predator-prey interactions, a circadian clock gene, a signal transduction pathway, and the JAK/STAT pathway.

This paper presents a comprehensive review and comparison of five computational methods for Bayesian model selection, based on MCMC simulations from posterior model parameter distributions. We apply these methods to a well-known and important class of models in financial time series analysis, namely GARCH and GARCH-t models for conditional return distributions (assuming normal and t-distributions). We compare their performance vis--vis the more common maximum likelihood-based model selection on both simulated and real market data. All five MCMC methods proved feasible in both cases, although differing in their computational demands. Results on simulated data show that for large degrees of freedom (where the t-distribution becomes more similar to a normal one), Bayesian model selection results in better decisions in favour of the true model than maximum likelihood. Results on market data show the feasibility of all model selection methods, mainly because the distributions appear to be decisively non-Gaussian. / Series: Report Series SFB "Adaptive Information Systems and Modelling in Economics and Management Science"

Gaussian processes are powerful nonparametric distributions over continuous functions that have become a standard tool in modern probabilistic machine learning. However, the applicability of Gaussian processes in the large-data regime and in hierarchical probabilistic models is severely limited by analytic and computational intractabilities. It is, therefore, important to develop practical approximate inference and learning algorithms that can address these challenges. To this end, this dissertation provides a comprehensive and unifying perspective of pseudo-point based deterministic approximate Bayesian learning for a wide variety of Gaussian process models, which connects previously disparate literature, greatly extends them and allows new state-of-the-art approximations to emerge. We start by building a posterior approximation framework based on Power-Expectation Propagation for Gaussian process regression and classification. This framework relies on a structured approximate Gaussian process posterior based on a small number of pseudo-points, which is judiciously chosen to summarise the actual data and enable tractable and efficient inference and hyperparameter learning. Many existing sparse approximations are recovered as special cases of this framework, and can now be understood as performing approximate posterior inference using a common approximate posterior. Critically, extensive empirical evidence suggests that new approximation methods arisen from this unifying perspective outperform existing approaches in many real-world regression and classification tasks. We explore the extensions of this framework to Gaussian process state space models, Gaussian process latent variable models and deep Gaussian processes, which also unify many recently developed approximation schemes for these models. Several mean-field and structured approximate posterior families for the hidden variables in these models are studied. We also discuss several methods for approximate uncertainty propagation in recurrent and deep architectures based on Gaussian projection, linearisation, and simple Monte Carlo. The benefit of the unified inference and learning frameworks for these models are illustrated in a variety of real-world state-space modelling and regression tasks.

This master's thesis deals with possible applications of Bayesian networks. The theoretical part is mainly of mathematical nature. At first, we focus on general probability theory and later we move on to the theory of Bayesian networks and discuss approaches to inference and to model learning while providing explanations of pros and cons of these techniques. The practical part focuses on applications that demand learning a Bayesian network, both in terms of network parameters as well as structure. These applications include general benchmarks, usage of Bayesian networks for knowledge discovery regarding the causes of criminality and exploration of the possibility of using a Bayesian network as a spam filter.

Shrinkage for time-varying parameter (TVP) models is investigated within a Bayesian framework, with the aim to automatically reduce time-varying Parameters to staticones, if the model is overfitting. This is achieved through placing the double gamma shrinkage prior on the process variances. An efficient Markov chain Monte Carlo scheme is devel- oped, exploiting boosting based on the ancillarity-sufficiency interweaving strategy. The method is applicable both to TVP models for univariate a swell as multivariate time series. Applications include a TVP generalized Phillips curve for EU area inflation modeling and a multivariate TVP Cholesky stochastic volatility model for joint modeling of the Returns from the DAX-30index.

The main topic of the paper is on-line filtering for non-Gaussian dynamic (state space) models by approximate computation of the first two posterior moments using efficient numerical integration. Based on approximating the prior of the state vector by a normal density, we prove that the posterior moments of the state vector are related to the posterior moments of the linear predictor in a simple way. For the linear predictor Gauss-Hermite integration is carried out with automatic reparametrization based on an approximate posterior mode filter. We illustrate how further topics in applied state space modelling such as estimating hyperparameters, computing model likelihoods and predictive residuals, are managed by integration-based Kalman-filtering. The methodology derived in the paper is applied to on-line monitoring of ecological time series and filtering for small count data. (author's abstract) / Series: Forschungsberichte / Institut für Statistik

The topic of the paper is filtering for non-Gaussian dynamic (state space) models by approximate computation of posterior moments using numerical integration. A Gauss-Hermite procedure is implemented based on the approximate posterior mode estimator and curvature recently proposed in 121. This integration-based filtering method will be illustrated by a dynamic trend model for non-Gaussian time series. Comparision of the proposed method with other approximations ([15], [2]) is carried out by simulation experiments for time series from Poisson, exponential and Gamma distributions. (author's abstract) / Series: Forschungsberichte / Institut für Statistik

This thesis considers the problem of performing inference on undirected graphical models with continuous state spaces. These models represent conditional independence structures that can appear in the context of Bayesian Machine Learning. In the thesis, we focus on computational methods and applications. The aim of the thesis is to demonstrate that the factorisation structure corresponding to the conditional independence structure present in high-dimensional models can be exploited to decrease the computational complexity of inference algorithms. First, we consider the smoothing problem on Hidden Markov Models (HMMs) and discuss novel algorithms that have sub-quadratic computational complexity in the number of particles used. We show they perform on par with existing state-of-the-art algorithms with a quadratic complexity. Further, a novel class of rejection free samplers for graphical models known as the Local Bouncy Particle Sampler (LBPS) is explored and applied on a very large instance of the Probabilistic Matrix Factorisation (PMF) problem. We show the method performs slightly better than Hamiltonian Monte Carlo methods (HMC). It is also the first such practical application of the method to a statistical model with hundreds of thousands of dimensions. In a second part of the thesis, we consider approximate Bayesian inference methods and in particular the Expectation Propagation (EP) algorithm. We show it can be applied as the backbone of a novel distributed Bayesian inference mechanism. Further, we discuss novel variants of the EP algorithms and show that a specific type of update mechanism, analogous to the mirror descent algorithm outperforms all existing variants and is robust to Monte Carlo noise. Lastly, we show that EP can be used to help the Particle Belief Propagation (PBP) algorithm in order to form cheap and adaptive proposals and significantly outperform classical PBP.

The multi-target tracking problem essentially involves the recursive joint estimation of the state of unknown and time-varying number of targets present in a tracking scene, given a series of observations. This problem becomes more challenging because the sequence of observations is noisy and can become corrupted due to miss-detections and false alarms/clutter. Additionally, the detected observations are indistinguishable from clutter. Furthermore, whether the target(s) of interest are point or extended (in terms of spatial extent) poses even more technical challenges. An approach known as random finite sets provides an elegant and rigorous framework for the handling of the multi-target tracking problem. With a random finite sets formulation, both the multi-target states and multi-target observations are modelled as finite set valued random variables, that is, random variables which are random in both the number of elements and the values of the elements themselves. Furthermore, compared to other approaches, the random finite sets approach possesses a desirable characteristic of being free of explicit data association prior to tracking. In addition, a framework is available for dealing with random finite sets and is known as finite sets statistics. In this thesis, advanced signal processing techniques are employed to provide enhancements to and develop new random finite sets based multi-target tracking algorithms for the tracking of both point and extended targets with the aim to improve tracking performance in cluttered environments. To this end, firstly, a new and efficient Kalman-gain aided sequential Monte Carlo probability hypothesis density (KG-SMC-PHD) filter and a cardinalised particle probability hypothesis density (KG-SMC-CPHD) filter are proposed. These filters employ the Kalman- gain approach during weight update to correct predicted particle states by minimising the mean square error between the estimated measurement and the actual measurement received at a given time in order to arrive at a more accurate posterior. This technique identifies and selects those particles belonging to a particular target from a given PHD for state correction during weight computation. The proposed SMC-CPHD filter provides a better estimate of the number of targets. Besides the improved tracking accuracy, fewer particles are required in the proposed approach. Simulation results confirm the improved tracking performance when evaluated with different measures. Secondly, the KG-SMC-(C)PHD filters are particle filter (PF) based and as with PFs, they require a process known as resampling to avoid the problem of degeneracy. This thesis proposes a new resampling scheme to address a problem with the systematic resampling method which causes a high tendency of resampling very low weight particles especially when a large number of resampled particles are required; which in turn affect state estimation. Thirdly, the KG-SMC-(C)PHD filters proposed in this thesis perform filtering and not tracking , that is, they provide only point estimates of target states but do not provide connected estimates of target trajectories from one time step to the next. A new post processing step using game theory as a solution to this filtering - tracking problem is proposed. This approach was named the GTDA method. This method was employed in the KG-SMC-(C)PHD filter as a post processing technique and was evaluated using both simulated and real data obtained using the NI-USRP software defined radio platform in a passive bi-static radar system. Lastly, a new technique for the joint tracking and labelling of multiple extended targets is proposed. To achieve multiple extended target tracking using this technique, models for the target measurement rate, kinematic component and target extension are defined and jointly propagated in time under the generalised labelled multi-Bernoulli (GLMB) filter framework. The GLMB filter is a random finite sets-based filter. In particular, a Poisson mixture variational Bayesian (PMVB) model is developed to simultaneously estimate the measurement rate of multiple extended targets and extended target extension was modelled using B-splines. The proposed method was evaluated with various performance metrics in order to demonstrate its effectiveness in tracking multiple extended targets.

Probabilistic models of decision making under various forms of uncertainty have been applied in recent years to numerous behavioral and model-based fMRI studies. These studies were highly successful in enabling a better understanding of behavior and delineating the functional properties of brain areas involved in decision making under uncertainty. However, as different studies considered different models of decision making under uncertainty, it is unclear which of these computational models provides the best account of the observed behavioral and neuroimaging data. This is an important issue, as not performing model comparison may tempt researchers to over-interpret results based on a single model. Here we describe how in practice one can compare different behavioral models and test the accuracy of model comparison and parameter estimation of Bayesian and maximum-likelihood based methods. We focus our analysis on two well-established hierarchical probabilistic models that aim at capturing the evolution of beliefs in changing environments: Hierarchical Gaussian Filters and Change Point Models. To our knowledge, these two, well-established models have never been compared on the same data. We demonstrate, using simulated behavioral experiments, that one can accurately disambiguate between these two models, and accurately infer free model parameters and hidden belief trajectories (e.g., posterior expectations, posterior uncertainties, and prediction errors) even when using noisy and highly correlated behavioral measurements. Importantly, we found several advantages of Bayesian inference and Bayesian model comparison compared to often-used Maximum-Likelihood schemes combined with the Bayesian Information Criterion. These results stress the relevance of Bayesian data analysis for model-based neuroimaging studies that investigate human decision making under uncertainty.