<p> Catalyzed by the invention of magnetic tape recording, audio production has transformed from technical to artistic, and the roles of producer, engineer, composer, and performer have merged for many forms of music. However, while these roles have changed, the way we interact with audio production tools has not and still relies on the conventions established in the 1970s for audio engineers. Users communicate their audio concepts to these complex tools using knobs and sliders that control low-level technical parameters. Musicians currently need technical knowledge of signals in addition to their musical knowledge to make novel music. However, many experienced and casual musicians simply do not have the time or desire to acquire this technical knowledge. While simpler tools (e.g. Apple's <i>GarageBand</i>) exist, they are limiting and frustrating to users. </p><p> To support these audio-production novices, we must build audio-production tools with affordances for them. We must identify interactions that enable novices to communicate their audio concepts without requiring technical knowledge and develop systems that can understand these interactions. </p><p> This dissertation advances our understanding of this problem by investigating three interaction types which are inspired by how novices communicate audio concepts to other people: <i>language, vocal imitation,</i> and <i> evaluation</i>. Because learning from an individual can be time consuming for a user, much of this dissertation focuses on how we can learn general audio concepts offline using crowdsourcing rather than user-specific audio concepts. This work introduces algorithms, frameworks, and software for learning audio concepts via these interactions and investigates the strengths and weaknesses of both the algorithms and the interaction types. These contributions provide a research foundation for a new generation of audio-production tools. </p><p> This problem is not limited to audio production tools. Other media production tools&mdash;such as software for graphics, image, and video design and editing&mdash;are also controlled by low-level technical parameters which require technical knowledge and experience to use effectively. The contributions in this dissertation to learn mappings from descriptive language and feedback to low-level control parameters may also be adapted for creative production tools in these other mediums. The contributions in this dissertation can unlock the creativity trapped in everyone who has the desire to make music and other media but does not have the technical skills required for today's tools.</p>

<p> This work seeks to define methodologies and techniques to analyze automotive fleet-tracking big data and provide sample results that have implications to the real world. To perform this work, vehicle fleet-tracking data from Odyne and Via Plug-in Hybrid Electric Trucks collected by the Electric Power Research Institute (EPRI) was used. Both CAN-communication bus signals and GPS data were recorded off of these vehicles with a second-by-second data collection rate. Colorado State University (CSU) was responsible for analyzing this data after it had been collected by EPRI and producing results with application to the real world. </p><p> A list of potential research questions is presented and an initial feasibility assessment is performed to determine how these questions might be answered using vehicle fleet-tracking data. Later, a subset of these questions are analyzed and answered in detail using the EPRI dataset. </p><p> The methodologies, techniques, and software used for this data analysis are described in detail. An algorithm that summarizes second-by-second vehicle tracking data into a list of higher-level driving and charging events is presented and utility factor (UF) curves and other statistics of interest are generated from this summarized event data. </p><p> In addition, another algorithm was built on the driving event identification algorithm to discretize the driving event data into approximately 90-second drive intervals. This allows for a regression model to be fit onto the data. A correlation between ambient temperature and equivalent vehicle fuel economy (in miles per gallon) is presented for Odyne and it is similar to the trend seen in conventional vehicle fuel economy vs. ambient temperature. It is also shown how ambient temperature variations can influence the vehicle fuel economy and there is a discussion about how changes in HVAC use could influence the fuel economy results. </p><p> It is also demonstrated how variations in the data analysis methodology can influence the final results. This provides evidence that vehicle fleet-tracking data analysis methodologies need to be defined to ensure that the data analysis results are of the highest quality. The questions and assumptions behind the presented analysis results are examined and a list of future work to address potential concerns and unanswered questions about the data analysis process is presented. Hopefully, this future work list will be beneficial to future vehicle data analysis projects. </p><p> The importance of using real-world driving data is demonstrated by comparing fuel economy results from our real-world data to the fuel economy calculated by EPA drive cycles. Utility factor curves calculated from the real-world data are also compared to standard utility factor curves that are presented in the SAE J2841 specification. Both of these comparisons showed a difference in real-world driving data, demonstrating the potential utility of evaluating vehicle technologies using the real-world big data techniques presented in this work. </p><p> Overall, this work documents some of the data analysis techniques that can be used for analyzing vehicle fleet-tracking big data and demonstrates the impact of the analysis results in the real world. It also provides evidence that the data analysis methodologies used to analyze vehicle fleet-tracking data need to be better defined and evaluated in future work.</p>

<p> Big data queries are increasing in complexity and the performance of data analytics is of growing importance. To this end, Big Data on high-performance computing (HPC) infrastructure is becoming a pathway to high-performance data analytics. The state of performance studies on this convergence between Big Data and HPC, however, is limited and ad hoc. A systematic performance study is thus timely and forms the core of this research. </p><p> This thesis investigates the challenges involved in developing Big Data applications with significant computations and strict latency guarantees on multicore HPC clusters. Three key areas it considers are thread models, affinity, and communication mechanisms. Thread models discuss the challenges of exploiting intra-node parallelism on modern multicore chips, while affinity looks at data locality and Non-Uniform Memory Access (NUMA) effects. Communication mechanisms investigate the difficulties of Big Data communications. For example, parallel machine learning depends on collective communications, unlike classic scientific simulations, which mostly use neighbor communications. Minimizing this cost while scaling out to higher parallelisms requires non-trivial optimizations, especially when using high-level languages such as Java or Scala. The investigation also includes a discussion on performance implications of different programming models such as dataflow and message passing used in Big Data analytics. The optimizations identified in this research are incorporated in developing the Scalable Parallel Interoperable Data Analytics Library (SPIDAL) in Java, which includes a collection of multidimensional scaling and clustering algorithms optimized to run on HPC clusters. </p><p> Besides presenting performance optimizations, this thesis explores a novel scheme for characterizing Big Data benchmarks. Fundamentally, a benchmark evaluates a certain performance-related aspect of a given system. For example, HPC benchmarks such as LINPACK and NAS Parallel Benchmark (NPB) evaluate the floating-point operations (flops) per second through a computational workload. The challenge with Big Data workloads is the diversity of their applications, which makes it impossible to classify them along a single dimension. Convergence Diamonds (CDs) is a multifaceted scheme that identifies four dimensions of Big Data workloads. These dimensions are problem architecture, execution, data source and style, and processing view. </p><p> The performance optimizations together with the richness of CDs provide a systematic guide to developing high-performance Big Data benchmarks, specifically targeting data analytics on large, multicore HPC clusters.</p>

<p> Medical informatics plays an important role in precision medicine, delivering the right information to the right person, at the right time. With the introduction and widespread adoption of electronic medical records, in the United States and world-wide, there is now a tremendous amount of health data available for analysis.</p><p> Electronic record phenotyping refers to the task of determining, from an electronic medical record entry, a concise descriptor of the patient, comprising of their medical history, current problems, presentation, etc. In inferring such a phenotype descriptor from the record, a computer, in a sense, "understands'' the relevant parts of the record. These phenotypes can then be used in downstream applications such as cohort selection for retrospective studies, real-time clinical decision support, contextual displays, intelligent search, and precise alerting mechanisms.</p><p> We are faced with three main challenges:</p><p> First, the unstructured and incomplete nature of the data recorded in the electronic medical records requires special attention. Relevant information can be missing or written in an obscure way that the computer does not understand. </p><p> Second, the scale of the data makes it important to develop efficient methods at all steps of the machine learning pipeline, including data collection and labeling, model learning and inference.</p><p> Third, large parts of medicine are well understood by health professionals. How do we combine the expert knowledge of specialists with the statistical insights from the electronic medical record?</p><p> Probabilistic graphical models such as Bayesian networks provide a useful abstraction for quantifying uncertainty and describing complex dependencies in data. Although significant progress has been made over the last decade on approximate inference algorithms and structure learning from complete data, learning models with incomplete data remains one of machine learning&rsquo;s most challenging problems. How can we model the effects of latent variables that are not directly observed?</p><p> The first part of the thesis presents two different structural conditions under which learning with latent variables is computationally tractable. The first is the "anchored'' condition, where every latent variable has at least one child that is not shared by any other parent. The second is the "singly-coupled'' condition, where every latent variable is connected to at least three children that satisfy conditional independence (possibly after transforming the data). </p><p> Variables that satisfy these conditions can be specified by an expert without requiring that the entire structure or its parameters be specified, allowing for effective use of human expertise and making room for statistical learning to do some of the heavy lifting. For both the anchored and singly-coupled conditions, practical algorithms are presented.</p><p> The second part of the thesis describes real-life applications using the anchored condition for electronic phenotyping. A human-in-the-loop learning system and a functioning emergency informatics system for real-time extraction of important clinical variables are described and evaluated.</p><p> The algorithms and discussion presented here were developed for the purpose of improving healthcare, but are much more widely applicable, dealing with the very basic questions of identifiability and learning models with latent variables - a problem that lies at the very heart of the natural and social sciences.</p>

<p> Vowel harmony has been of great interest in phonological research. It has been widely accepted that vowel harmony is a phonetically natural phenomenon, which means that it is a common pattern because it provides advantages to the speaker in articulation and to the listener in perception. </p><p> Exceptional patterns proved to be a challenge to the phonetically grounded analysis as they, by their nature, introduce phonetically disadvantageous sequences to the surface form, that consist of harmonically different vowels. Such forms are found, for example in the Finnish stem <i>tuoli</i> 'chair' or in the Hungarian suffixed form <i><b>hi:d-hoz</b></i> 'to the bridge', both word forms containing a mix of front and back vowels. There has recently been evidence shown that there might be a phonetic level explanation for some exceptional patterns, as the possibility that some vowels participating in irregular stems (like the vowel [i] in the Hungarian stem <i> hi:d</i> 'bridge' above) differ in some small phonetic detail from vowels in regular stems. The main question has not been raised, though: does this phonetic detail matter for speakers? Would they use these minor differences when they have to categorize a new word as regular or irregular?</p><p> A different recent trend in explaining morphophonological exceptionality by looking at the phonotactic regularities characteristic of classes of stems based on their morphological behavior. Studies have shown that speakers are aware of these regularities, and use them as cues when they have to decide what class a novel stem belongs to. These sublexical phonotactic regularities have already been shown to be present in some exceptional patterns vowel harmony, but many questions remain open: how is learning the static generalization linked to learning the allomorph selection facet of vowel harmony? How much does the effect of consonants on vowel harmony matter, when compared to the effect of vowel-to-vowel correspondences?</p><p> This dissertation aims to test these two ideas --- that speakers use phonetic cues and/or that they use sublexical phonotactic regularities in categorizing stems as regular or irregular --- and attempt to answer the more detailed questions, like the effect of consonantal patterns on exceptional patterns or the link between allomorph selection and static phonotactic generalizations as well. The phonetic hypothesis is tested on the Hungarian antiharmonicity pattern (stems with front vowels consistently selecting back suffixes, like in the example <i>hi:d-hoz</i> 'to the bridge' above), and the results indicate that while there may be some small phonetic differences between vowels in regular and irregular stems, speakers do not use these, or even enhanced differences when they have to categorize stems.</p><p> The sublexical hypothesis is tested and confirmed by looking at the disharmonicity pattern in Finnish. In Finnish, stems that contain both back and certain front vowels are frequent and perfectly grammatical, like in the example <i> tuoli</i> 'chair' above, while the mixing of back and some other front vowels is very rare and mostly confined to loanwords like <i>amat&oslash;&oslash;ri </i> 'amateur'. It will be seen that speakers do use sublexical phonotactic regularities to decide on the acceptability of novel stems, but certain patterns that are phonetically or phonologically more natural (vowel-to-vowel correspondences) seem to matter much more than other effects (like consonantal effects).</p><p> Finally, a computational account will be given on how exceptionality might be learned by speakers by using maximum entropy grammars available in the literature to simulate the acquisition of the Finnish disharmonicity pattern. It will be shown that in order to clearly model the overall behavior on the exact pattern, the learner has to have access not only to the lexicon, but also to the allomorph selection patterns in the language.</p>

<p> Introduced in 2006, Deep Learning has made large strides in both supervised an unsupervised learning. The abilities of Deep Learning have been shown to beat both generic and highly specialized classification and clustering techniques with little change to the underlying concept of a multi-layer perceptron. Though this has caused a resurgence of interest in neural networks, many of the drawbacks and pitfalls of such systems have yet to be addressed after nearly 30 years: speed of training, local minima and manual testing of hyper-parameters.</p><p> In this thesis we propose using an evolutionary technique in order to work toward solving these issues and increase the overall quality and abilities of Deep Learning Networks. In the evolution of a population of autoencoders for input reconstruction, we are able to abstract multiple features for each autoencoder in the form of hidden nodes, scoring the autoencoders based on their ability to reconstruct their input, and finally selecting autoencoders for crossover and mutation with hidden nodes as the chromosome. In this way we are able to not only quickly find optimal abstracted feature sets but also optimize the structure of the autoencoder to match the features being selected. This also allows us to experiment with different training methods in respect to data partitioning and selection, reducing overall training time drastically for large and complex datasets. This proposed method allows even large datasets to be trained quickly and efficiently with little manual parameter choice required by the user, leading to faster, more accurate creation of Deep Learning Networks.</p>

The registration of 3D preoperative medical data to patients is a key task in developing computer assisted surgery systems. In computer assisted surgery, the patient in the operation theatre must be aligned with the coordinate system in which the preoperative data has been acquired, so that the planned surgery based on the preoperative data can be carried out under the guidance of the computer assisted surgery system. The aim of this research is to investigate registration algorithms for developing computer assisted bone surgery systems. We start with reference mark registration. New interpretations are given to the development of well knowm algorithms based on singular value decomposition, polar decomposition techniques and the unit quaternion representation of the rotation matrix. In addition, a new algorithm is developed based on the estimate of the rotation axis. For non-land mark registration, we first develop iterative closest line segment and iterative closest triangle patch registrations, similar to the well known iterative closest point registration, when the preoperative data are dense enough. We then move to the situation where the preoperative data are not dense enough. Implicit fitting is considered to interpolate the gaps between the data . A new ellipsoid fitting algorithm and a new constructive implicit fitting strategy are developed. Finally, a region to region matching procedure is proposed based on our novel constructive implicit fitting technique. Experiments demonstrate that the new algorithm is very stable and very efficient.

Over the past decade, Genetic Programming (GP) has been the subject of a significant amount of research, but this has resulted in the solution of few complex real -world problems. In this work, I propose that, for some relatively simple, non safety -critical embedded control applications, GP can be used as a practical alternative to software developed by humans. Embedded control software has become a branch of software engineering with distinct temporal, interface and resource constraints and requirements. This results in a characteristic software structure, and by examining this, the effective decomposition of an overall problem into a number of smaller, simpler problems is performed. It is this type of problem amelioration that is suggested as a method whereby certain real -world problems may be rendered into a soluble form suitable for GP. In the course of this research, the body of published GP literature was examined and the most important changes to the original GP technique of Koza are noted; particular focus is made upon GP techniques involving an element of concurrency -which is central to this work. This search highlighted few applications of GP for the creation of software for complex, real -world problems -this was especially true in the case of multi thread, multi output solutions. To demonstrate this Idea, a concurrent Linear GP (LGP) system was built that creates a multiple input -multiple output solution using a custom low -level evolutionary language set, combining both continuous and Boolean data types. The system uses a multi -tasking model to evolve and execute the required LGP code for each system output using separate populations: Two example problems -a simple fridge controller and a more complex washing machine controller are described, and the problems encountered and overcome during the successful solution of these problems, are detailed. The operation of the complete, evolved washing machine controller is simulated using a graphical LabVIEWapplication. The aim of this research is to propose a general purpose system for the automatic creation of control software for use in a range of problems from the target problem class -without requiring any system tuning: In order to assess the system search performance sensitivity, experiments were performed using various population and LGP string sizes; the experimental data collected was also used to examine the utility of abandoning stalled searches and restarting. This work is significant because it identifies a realistic application of GP that can ease the burden of finite human software design resources, whilst capitalising on accelerating computing potential.

The limitations of speech output technology emphasise the need for exploratory psychological research to maximise the effectiveness of speech as a display medium in human-computer interaction. Stage 1 of this study reviewed speech implementation research, focusing on general issues for tasks, users and environments. An analysis of design issues was conducted, related to the differing methodologies for synthesised and digitised message production. A selection of ergonomic guidelines were developed to enhance effective speech interface design. Stage 2 addressed the negative reactions of users to synthetic speech in spite of elegant dialogue structure and appropriate functional assignment. Synthetic speech interfaces have been consistently rejected by their users in a wide variety of application domains because of their poor quality. Indeed the literature repeatedly emphasises quality as being the most important contributor to implementation acceptance. In order to investigate this, a converging operations approach was adopted. This consisted of a series of five experiments (and associated pilot studies) which homed in on the specific characteristics of synthetic speech that determine the listeners varying perceptions of its qualities, and how these might be manipulated to improve its aesthetics. A flexible and reliable ratings interface was designed to display DECtalk speech variations and record listeners perceptions. In experiment one, 40 participants used this to evaluate synthetic speech variations on a wide range of perceptual scales. Factor analysis revealed two main factors: "listenability" accounting for 44.7% of the variance and correlating with the DECtalk "smoothness" parameter to . 57 (p<0.005) and "richness" to . 53 (p<0.005); "assurance" accounting for 12.6% of the variance and correlating with "average pitch" to . 42 (p<0.005) and "head size" to. 42 (p<0.005). Complimentary experiments were then required in order to address appropriate voice design for enhanced listenability and assurance perceptions. With a standard male voice set, 20 participants rated enhanced smoothness and attenuated richness as contributing significantly to speech listenability (p<0.001). Experiment three using a female voice set yielded comparable results, suggesting that further refinements of the technique were necessary in order to develop an effective methodology for speech quality optimization. At this stage it became essential to focus directly on the parameter modifications that are associated with the the aesthetically pleasing characteristics of synthetic speech. If a reliable technique could be developed to enhance perceived speech quality, then synthesis systems based on the commonly used DECtalk model might assume some of their considerable yet unfulfilled potential. In experiment four, 20 subjects rated a wide range of voices modified across the two main parameters associated with perceived listenability, smoothness and richness. The results clearly revealed a linear relationship between enhanced smoothness and attenuated richness and significant improvements in perceived listenability (p<0.001 in both cases). Planned comparisons conducted were between the different levels of the parameters and revealed significant listenability enhancements as smoothness was increased, and a similar pattern as richness decreased. Statistical analysis also revealed a significant interaction between the two parameters (p<0.001) and a more comprehensive picture was constructed. In order to expand the focus of and enhance the generality of the research, it was now necessary to assess the effects of synthetic speech modifications whilst subjects were undertaking a more realistic task. Passively rating the voices independent of processing for meaning is arguably an artificial task which rarely, if ever, would occur in 'real-world' settings. In order to investigate perceived quality in a more realistic task scenario, experiment five introduced two levels of information processing load. The purpose of this experiment was firstly to see if a comprehension load modified the pattern of listenability enhancements, and secondly to see if that pattern differed between high and and low load. Techniques for introducing cognitive load were investigated and comprehension load was selected as the most appropriate method in this case. A pilot study distinguished two levels of comprehension load from a set of 150 true/false sentences and these were recorded across the full range of parameter modifications. Twenty subjects then rated the voices using the established listenability scales as before but also performing the additional task of processing each spoken stimuli for meaning and determining the authenticity of the statements. Results indicated that listenability enhancements did indeed occur at both levels of processing although at the higher level variations in the pattern occured. A significant difference was revealed between optimal parameter modifications for conditions of high and low cognitive load (p<0.05). The results showed that subjects perceived the synthetic voices in the high cognitive load condition to be significantly less listenable than those same voices in the low cognitive load condition. The analysis also revealed that this effect was independent of the number of errors made. This result may be of general value because conclusions drawn from this findings are independent of any particular parameter modifications that may be exclusively available to DECtalk users. Overall, the study presents a detailed analysis of the research domain combined with a systematic experimental program of synthetic speech quality assessment. The experiments reported establish a reliable and replicable procedure for optimising the aesthetically pleasing characteristics of DECtalk speech, but the implications of the research extend beyond the boundaries of a particular synthesiser. Results from the experimental program lead to a number of conclusions, the most salient being that not only does the synthetic speech designer have to overcome the general rejection of synthetic voices based on their poor quality by sophisticated customisation of synthetic voice parameters, but that he or she needs to take into account the cognitive load of the task being undertaken. The interaction between cognitive load and optimal settings for synthesis requires direct consideration if synthetic speech systems are going to realise and maximise their potential in human computer interaction.

The realism of character animation is associated with a number of tasks ranging from modelling, skin defonnation, motion generation to rendering. In this research we are concerned with two of them: skeletonization and weight assignment for skin deformation. The fonner is to generate a skeleton, which is placed within the character model and links the motion data to the skin shape of the character. The latter assists the modelling of realistic skin shape when a character is in motion. In the current animation production practice, the task of skeletonization is primarily undertaken by hand, i.e. the animator produces an appropriate skeleton and binds it with the skin model of a character. This is inevitably very time-consuming and costs a lot of labour. In order to improve this issue, in this thesis we present an automatic skeletonization framework. It aims at producing high-quality animatible skeletons without heavy human involvement while allowing the animator to maintain the overall control of the process. In the literature, the tenn skeletonization can have different meanings. Most existing research on skeletonization is in the remit of CAD (Computer Aided Design). Although existing research is of significant reference value to animation, their downside is the skeleton generated is either not appropriate for the particular needs of animation, or the methods are computationally expensive. Although some purpose-build animation skeleton generation techniques exist, unfortunately they rely on complicated post-processing procedures, such as thinning and pruning, which again can be undesirable. The proposed skeletonization framework makes use of a new geometric entity known as the 3D silhouette that is an ordinary silhouette with its depth information recorded. We extract a curve skeleton from two 3D silhouettes of a character detected from its two perpendicular projections. The skeletal joints are identified by down sampling the curve skeleton, leading to the generation of the final animation skeleton. The efficiency and quality are major performance indicators in animation skeleton generation. Our framework achieves the former by providing a 2D solution to the 3D skeletonization problem. Reducing in dimensions brings much faster performances. Experiments and comparisons are carried out to demonstrate the computational simplicity. Its accuracy is also verified via these experiments and comparisons. To link a skeleton to the skin, accordingly we present a skin attachment framework aiming at automatic and reasonable weight distribution. It differs from the conventional algorithms in taking topological information into account during weight computation. An effective range is defined for a joint. Skin vertices located outside the effective range will not be affected by this joint. By this means, we provide a solution to remove the influence of a topologically distant, hence highly likely irrelevant joint on a vertex. A user-defined parameter is also provided in this algorithm, which allows different deformation effects to be obtained according to user's needs. Experiments and comparisons prove that the presented framework results in weight distribution of good quality. Thus it frees animators from tedious manual weight editing. Furthermore, it is flexible to be used with various deformation algorithms.