Real-time realistic image synthesis for virtual environments has been one of the most actively researched areas in computer graphics for over a decade. Images that display physically correct illumination of an environment can be simulated by evaluating a multi-dimensional integral equation, called the rendering equation, over the surfaces of the environment. Many global illumination algorithms such as pathtracing, photon mapping and distributed ray-tracing can produce realistic images but are generally unable to cope with dynamic lighting and objects at interactive rates. It still remains one of most challenging problems to simulate physically correctly illuminated dynamic environments without a substantial preprocessing step. In this thesis we present a rendering system for dynamic environments by implementing a customized rasterizer for global illumination entirely on the graphics hardware, the Graphical Processing Unit. Our research focuses on a parameterization of discrete visibility field for efficient indirect illumination computation. In order to generate the visibility field, we propose a CUDA-based (Compute Unified Device Architecture) rasterizer which builds Layered Hit Buffers (LHB) by rasterizing polygons into multi-layered structural buffers in parallel. The LHB provides a fast visibility function for any direction at any point. We propose a cone approximation solution to resolve an aliasing problem due to limited directional discretization. We also demonstrate how to remove structure noises by adapting an interleaved sampling scheme and discontinuity buffer. We show that a gathering method amortized with a multi-level Quasi Mont Carlo method can evaluate the rendering equation in real-time. The method can realize real-time walk-through of a complex virtual environment that has a mixture of diffuse and glossy reflection, computing multiple indirect bounces on the fly. We show that our method is capable of simulating fully dynamic environments including changes of view, materials, lighting and objects at interactive rates on commodity level graphics hardware.

A large proportion of Machine Learning (ML) research focuses on designing algorithms that require minimal input from the human. However, ML algo- rithms are now widely used in various areas of engineering to design and build systems that interact with the human user and thus need to “learn” from this interaction. In this work, we concentrate on algorithms that learn from user interaction. A significant part of the dissertation is devoted to learning in the bandit setting. We propose a general framework for handling dependencies across arms, based on the new assumption that the mean-reward function is drawn from a Gaussian Process. Additionally, we propose an alternative method for arm selection using Thompson sampling and we apply the new algorithms to a grammar learning problem. In the remainder of the dissertation, we consider content-based image re- trieval in the case when the user is unable to specify the required content through tags or other image properties and so the system must extract infor- mation from the user through limited feedback. We present a novel Bayesian approach that uses latent random variables to model the systems imperfect knowledge about the users expected response to the images. An impor- tant aspect of the algorithm is the incorporation of an explicit exploration- exploitation strategy in the image sampling process. A second aspect of our algorithm is the way in which its knowledge of the target image is updated given user feedback. We considered a few algorithms to do so: variational Bayes, Gibbs sampling and a simple uniform update. We show in experi- ments that the simple uniform update performs best. The reason is because, unlike the uniform update, both variational Bayes and Gibbs sampling tend to focus on a small set of images aggressively.

Recent advances in virtual reality, 3d computer generated graphics and computer vision are making the goal of producing a compelling interactive 3d face to face communication system more tractable. The problem with producing such a system is reconstructing the 3d geometry of the users in real-time. There are many ways of tackling this problem however many of them require prior knowledge (i.e model fitting methods). These add unnecessary constraints and limit the usability of the system to reconstructing known entities. Other high quality methods using laser triangulation require too many samples and therefore cannot handle dynamic and deformable shapes such as the human face. A more suited approach is to use stereo based algorithm that function using two of more views and augmenting their capabilities using structured light. The work presented in this thesis will examine and evaluate various stereo vision algorithms and hybrids with the goal of producing accurate 3d representations of human faces in real time. Various dynamic programming algorithms will be presented and hybrid variations. These will be extended into the space-time domain and the impact of using different structured light patterns with various algorithms and cost functions will be examined. Most real-time correspondence algorithms are limited to producing pixel value disparities; these can be augmented into producing sub-pixel disparities by smoothing functions. Applying such smoothing functions tends to remove detail. Another approach is to use non-linear optimization on a spatial-temporal warp function. These algorithms tend to be very computationally expensive and therefore not feasible for real time applications. With recent development of GPUs (Graphics Processing Units) driven by the consumer demand for complex real time 3d graphics, these cards are capable of processing large amounts of data in parallel. This makes them very amenable to solving large linear algebra problems. . The result being a tuneable stereo reconstruction framework that has been reformulated into streaming problems in order to be processed on the GPU to produce real time sub-pixel depth maps of human faces that can be triangulated to produce accurate 3d models.

Many researchers has conjectured, argued, or in some cases demonstrated, that bio-plausibility can bring about emergent properties such as adaptability, scalability, fault-tolerance, self-repair, reliability, and autonomy to bio-inspired intelligent systems. Evolutionary-developmental (evo-devo) spiking neural networks are a very bio-plausible mixture of such bio-inspired intelligent systems that have been proposed and studied by a few researchers. However, the general trend is that the complexity and thus the computational cost grow with the bio-plausibility of the system. FPGAs (Field- Programmable Gate Arrays) have been used and proved to be one of the flexible and cost efficient hardware platforms for research' and development of such evo-devo systems. However, mapping a bio-plausible evo-devo spiking neural network to an FPGA is a daunting task full of different constraints and trade-offs that makes it, if not infeasible, very challenging. This thesis explores the challenges, trade-offs, constraints, practical issues, and some possible approaches in achieving bio-plausibility in creating evolutionary developmental spiking neural microcircuits in an FPGA through a practical investigation along with a series of case studies. In this study, the system performance, cost, reliability, scalability, availability, and design and testing time and complexity are defined as measures for feasibility of a system and structural accuracy and consistency with the current knowledge in biology as measures for bio-plausibility. Investigation of the challenges starts with the hardware platform selection and then neuron, cortex, and evo-devo models and integration of these models into a whole bio-inspired intelligent system are examined one by one. For further practical investigation, a new PLAQIF Digital Neuron model, a novel Cortex model, and a new multicellular LGRN evo-devo model are designed, implemented and tested as case studies. Results and their implications for the researchers, designers of such systems, and FPGA manufacturers are discussed and concluded in form of general trends, trade-offs, suggestions, and recommendations.

Cancer is a disease in which the controls that usually ensure the coordinated behaviour of individual cells break down. This rarely happens all at once. Instead, the clone of cells that grows into a developing tumour is under high selection pressure, leading to the evolution of a complex and diverse population of related cells that have accumulated a wide range of genetic defects. One of the most evident but poorly characterized of these genetic abnormalities is a disorder in the number of chromosomes, or aneuploidy. Aneuploidy can arise though several different mechanisms. The project explores one such mechanism - chromosome missegregation during cell division- and its role in oncogenesis. To address the role that chromosome missegregation may have in the development of cancer a computational model was devised. We then defined the behaviour of individual cells, their genomes and a tissue niche, which could be used in simulations to explore the different types of cell behaviour likely to arise as the result of chromosome missegregation. This model was then used to better understand how defects in chromosome segregation affect cancer development and tumour evolution during cancer therapy. In stochastic simulations, chromosome missegregation events at cell division lead to the generation of a diverse population of aneuploid clones that over time exhibit hyperplastic growth. Significantly, the course of cancer evolution depends on genetic linkage, as the structure of chromosomes lost or gained through missegregation events and the level of genetic instability function in tandem to determine whether tumour growth is driven primarily by the loss of tumour suppressors or by the overexpression of oncogenes. As a result, simulated cancers diff er in their level of genetic stability and in their growth rates. We then used this system to investigate the consequences of these differences in tumour heterogeneity for anti¬cancer therapies based on surgery and anti-mitotic drugs that selectively target proliferating cells. Results show that simulated treatments induce a transient delay in tumour growth, and reveal a significant difference in the efficacy of different therapy regimes in treating genetically stable and unstable tumours. These data support clinical observations in which a poor prognosis is correlated with a high level of chromosome missegregation. However, simulations run in parallel also exhibit a wide range of behaviours, and the response of individual simulations (equivalent to single tumours) to anti-cancer therapy prove extremely variable. The model therefore highlights the difficulties of predicting the outcome of a given anti-cancer treatment, even in cases in which it is possible to determine the genotype of the entire set of cells within the developing tumour.

We investigate the use of rejection applied to supervised learning for predicting the price direction of five foreign exchange currencies. We present two novel models which specifically take into account the random walk hypothesis when learning and predicting financial datasets. Both models project and then search a feature space for patterns and neighbourhoods unlikely to have arisen from a random process. The models invoke the human reply to an unfamiliar question of ‘I don’t know’ by rejecting (ignoring) training and/or test samples which do not satisfy checks for spurious relationships. The novel algorithms within this thesis are shown to significantly improve on both forecasting accuracy and economic viability when compared to several supervised learning reject and non-reject algorithms - the k-nearest neighbour and support vector machine algorithms are the main source of comparison. Reject-based models in general are shown to improve on the non-reject methods. Furthermore, several other contributions are noted within this thesis, namely: i) introducing intra-day data for forecasting daily price changes improves accuracy, ii) reducing the size of the time steps from one day to five minutes increased accuracy across all models; iii) forecasting accuracy was nearly always shown to reduce, across all models, after the events of the credit crisis (the years 2007 and 2009 are compared).

There is no point expending time and effort developing a model if it is based on data that is out of date. Many models require large amounts of data from a variety of heterogeneous sources. This data is subject to frequent and unannounced changes. It may only be possible to know that data has fallen out of date by reconstructing the model with the new data but this leads to further problems. How and when does the data change and when does the model need to be rebuilt? At best, the model will need to be continually rebuilt in a desperate attempt to remain current. At worst, the model will be producing erroneous results. The recent advent of automated and semi-automated data-processing and analysis tools in the biological sciences has brought about a rapid expansion of publicly available data. Many problems arise in the attempt to deal with this magnitude of data; some have received more attention than others. One significant problem is that data within these publicly available databases is subject to change in an unannounced and unpredictable manner. Large amounts of complex data from multiple, heterogeneous sources are obtained and integrated using a variety of tools. These data and tools are also subject to frequent change, much like the biological data. Reconciling these changes, coupled with the interdisciplinary nature of in silico biological experimentation, presents a significant problem. We present the ExperimentBuilder, an application that records both the current and previous states of an experimental environment. Both the data and metadata about an experiment are recorded. The current and previous versions of each of these experimental components are maintained within the ExperimentBuilder. When any one of these components change, the ExperimentBuilder estimates not only the impact within that specific experiment, but also traces the impact throughout the entire experimental environment. This is achieved with the use of keyword profiles, a heuristic tool for estimating the content of the experimental component. We can compare one experimental component to another regardless of their type and content and build a network of inter-component relationships for the entire environment. Ultimately, we can present the impact of an update as a complete cost to the entire environment in order to make an informed decision about whether to recalculate our results.

Decisions about which features to include in a new system or the next release of an existing one are critical to the success of software products. Such decisions should be informed by the needs of the users and stakeholders. But how can we make such decisions when the number of potential features and the number of individual stakeholders are very large? This problem is particularly important when stakeholders’ needs are gathered online through the use of discussion forums and web-based feature request management systems. Existing requirements decision-making techniques are not adequate in this context because they do not scale well to such large numbers of feature requests or stakeholders. This thesis addresses this problem by presenting and evaluating clustering methods to facilitate requirements selection and optimization when requirements preferences are elicited from a very large number of stakeholders. Firstly, it presents a novel method for identifying groups of stakeholders with similar preferences for requirements. It computes the representative preferences for the resulting groups and provides additional insights in trends and divergences in stakeholders’ preferences which may be used to aid the decision making process. Secondly, it presents a method to help decision-makers identify key similarities and differences among large sets of optimal design decisions. The benefits of these techniques are demonstrated on two real-life projects - one concerned with selecting features for mobile phones and the other concerned with selecting requirements for a rights and access management system.

Latent linear models are core to much of machine learning and statistics. Specific examples of this model class include Bayesian generalised linear models, Gaussian process regression models and unsupervised latent linear models such as factor analysis and principal components analysis. In general, exact inference in this model class is computationally and analytically intractable. Approximations are thus required. In this thesis we consider deterministic approximate inference methods based on minimising the Kullback-Leibler (KL) divergence between a given target density and an approximating `variational' density. First we consider Gaussian KL (G-KL) approximate inference methods where the approximating variational density is a multivariate Gaussian. Regarding this procedure we make a number of novel contributions: sufficient conditions for which the G-KL objective is differentiable and convex are described, constrained parameterisations of Gaussian covariance that make G-KL methods fast and scalable are presented, the G-KL lower-bound to the target density's normalisation constant is proven to dominate those provided by local variational bounding methods. We also discuss complexity and model applicability issues of G-KL and other Gaussian approximate inference methods. To numerically validate our approach we present results comparing the performance of G-KL and other deterministic Gaussian approximate inference methods across a range of latent linear model inference problems. Second we present a new method to perform KL variational inference for a broad class of approximating variational densities. Specifically, we construct the variational density as an affine transformation of independently distributed latent random variables. The method we develop extends the known class of tractable variational approximations for which the KL divergence can be computed and optimised and enables more accurate approximations of non-Gaussian target densities to be obtained.

We consider the problem of optimally allocating a limited budget to acquire relevance judgments when constructing an information retrieval test collection. We assume that there is a large set of test queries, for each of which a large number of documents need to be judged. However, the available budget only permits to judge a subset of them. We begin by developing a mathematical framework for query selection as a mechanism for reducing the cost of constructing information retrieval test collections. The mathematical framework provides valuable insights into properties of the optimal subset of queries. These are that the optimal subset of queries should be least correlated with one another, but have a strong correlation with the rest of queries. In contrast to previous work, which is mostly retrospective, our mathematical framework does not assume that relevance judgments are available a priori, and hence is designed to work in practice. The mathematical framework is then extended to accommodate both the query selection and document selection approaches to arrive at a unified budget allocation method that prioritizes query-document pairs and selects a subset of them with the highest priority scores to be judged. The unified budget allocation is formulated as a convex optimization, thereby permitting efficient solution and providing a flexible framework to incorporate various optimization constraints. Once a subset of query-document pairs are selected, crowdsourcing can be used to collect associated relevance judgments. While the labels provided by crowdsourcing are relatively inexpensive, they vary in quality, introducing noise into the relevance judgments. To deal with noisy relevance judgments, multiple labels for a document are collected from different assessors. It is common practice in information retrieval to use majority voting to aggregate multiple labels. In contrast, we develop a probabilistic model that provides accurate relevance judgments with a smaller number of labels collected per document. We demonstrate the effectiveness of our cost optimization approach on three experimental data, namely: (i) various TREC tracks, (ii) a web test collection of an online search engine, and (iii) crowdsourced data collected for the INEX 2010 Book Search track. Our approach should assist research institutes, e.g. National Institute and Standard Technology (NIST), and commercial search engines, e.g. Google and Bing, to construct test collections where there are large document collections and large query logs, but where economic constraints prohibit gathering comprehensive relevance judgments.