Search

Deep learning is experiencing a revolution with tremendous progress because of the availability of large datasets and computing resources. The development of deeper and larger neural network models has made significant progress recently in boosting the accuracy of many applications, such as image classification, image captioning, object detection, and language translation. However, despite the opportunities they offer, existing deep learning approaches are impractical for many applications due to the following challenges. Many applications exist with only limited amounts of annotated training data, or the collected labelled training data is too expensive. Such scenarios impose significant drawbacks for deep learning methods, which are not designed for limited data and suffer from performance decay. Especially for generative tasks, because the data for many generative tasks is difficult to obtain from the real world and the results they generate are difficult to control. As deep learning algorithms become more complicated increasing the workload for researchers to train neural network models and manage the life-cycle deep learning workflows, including the model, dataset, and training pipeline, the demand for efficient deep learning development is rising. Practical deep learning should achieve adequate performance from the limited training data as well as be based on efficient deep learning development processes. In this thesis, we propose several novel methods to improve the practicability of deep generative models and development processes, leading to four contributions. First, we improve the visual quality of synthesising images conditioned on text descriptions without requiring more manual labelled data, which provides controllable generated results using object attribute information from text descriptions. Second, we achieve unsupervised image-to-image translation that synthesises images conditioned on input images without requiring paired images to supervise the training, which provides controllable generated results using semantic visual information from input images. Third, we deliver semantic image synthesis that synthesises images conditioned on both image and text descriptions without requiring ground truth images to supervise the training, which provides controllable generated results using both semantic visual and object attribute information. Fourth, we develop a research-oriented deep learning library called TensorLayer to reduce the workload of researchers for defining models, implementing new layers, and managing the deep learning workflow comprised of the dataset, model, and training pipeline. In 2017, this library has won the best open source software award issued by ACM Multimedia (MM).

This research generates insights into the rapid multiplication of agricultural co-operatives in Mainland China before and after the introduction of China's first law on agricultural co-operatives in 2007. More specifically, the research utilises a multiple-case research method to explore the specific context of small-scale tea co-operatives and their supply chains. The research has several objectives. First, to explore how the values of co-operatives conform to Chinese society and culture. Second, to develop a typology of Chinese co-operatives based upon data collected for the tea co-operative industry. Third, to analyse the relationships among the membership of different types of co-operatives to determine if they result in different operational performance. Fourth, to explore how internal factors affect the power dynamics between a co-operative and its stakeholders. The data informing this research was generated through semi-structured qualitative interviews. Triangulation is achieved through the analysis of documentary sources and the researcher's observations (including participation in tea production). There are three principal findings resulting from the empirical research relating to: the recognition and comparison of Chinese co-operatives and Western co-operative values; building a construct model of Chinese agricultural co-operatives with propositions derived from embeddedness and institutional theories; and the creation of two typologies describing contemporary agricultural co-operatives in China.

This thesis consists of three chapters, which are works during my PhD study. In the first two chapters, I investigate the estimation of nonparametric and semiparametric econometric models widely used in empirical studies when the data is mismeasured. In the last chapter, my attention moves to the estimation of the effects of the social interactions. In Chapter 1, I study the estimation of the nonparametric additive model in the presence of a mismeasured covariate. In such a situation, the conventional method may cause severe bias. Therefore, I propose a new estimator. The estimation procedure is divided into two stages. In the first stage, to adept to the additive structure, I use a series method. And to deal with the ill-posedness brought by the mismeasurement, I introduce a ridge parameter. The convergence rate is then derived for the first stage estimator. For the distributional results required for inference, based on the first stage estimator, I implement the one-step back-fitting with a deconvolution kernel. Asymptotic normality is derived for the second stage estimator. Chapter 2 investigates the sharp regression-discontinuity (SRD) design when there is a continuously distributed measurement error in the running variable. In such a situation, the discontinuity at the cut-off disappear completely, and using the conventional SRD method cause severe bias. To overcome this, I develop a new estimator of the average treatment effect at the cut-off. Two separate cases characterized by the observability of the treatment status are considered. In the case of observed treatment status, the proposed estimator is the difference between the deconvolution local linear estimators based on treated and control groups. In the case of unobserved treatment status, the observed running variable cannot be used to divide the sample due to the presence of measurement errors. So, the one-sided kernel functions are implemented, and an additional ridging parameter is introduced for regularization. Asymptotic properties of proposed estimators are derived for both cases. Chapter 3 develops a new method to estimate the spillover effects using the factor structure of the variables generating the spillovers. Specifically, we find that such a factor structure implies constraints on the spillovers, which can be utilized to improve the performance of the existing estimator, like LASSO, by adding the factor-induced constraints. The L2 error bound is derived for the proposed estimator. Compared with the unconstrained case, the proposed estimator is more accurate in the sense that it has approximately sharper error bound. Simulation results demonstrate our findings.

Osseointegration trans-femoral implants are a novel orthopaedic anchoring method for connecting the stump and prosthesis after an above-knee amputation operation. The implant’s long-term stability is considered to be the essential condition of the osseointegration operation and rehabilitation. However, the stability could be diminished by the bone remodelling that occurs over. No such bone remodelling has been investigated in patients with osseointegrated trans-femoral implants. A bone remodelling algorithm has been developed with computer aided simulation by applying the Finite Element Method (FEM) for patients with osseointegrated trans-femoral implant. These bone remodelling algorithms predict the change of bone geometry that occurs in these patients. Various mechanical factors have been considered as the stimulus that governs the bone growth and bone loss. To achieve the objective of deducing the bone remodelling algorithm, the feasibility and the reliability of the modelling method for Finite Element (FE) femur models were first investigated. A technique was developed for creating FE models of the non-implanted and implanted femur. This has enabled the effect of using different size implants, where clinically representative incomplete contact between femur and implant are modelled, to be assessed. Subsequent FE modelling was undertaken on implanted femurs where bone remodelling has occurred over a 5 year period using long term X-ray data to update the original femur models. This part of study led to the conclusion that the non-adapted and the adapted femur models can be accurately reconstructed using the per-surgery computer tomography (CT) scans and long-term X-ray images. The results of the Finite Element Analysis (FEA) from the femur models showed the bone remodelling rate was regulated by a double control mechanism and a bone remodelling equation with two mechanical variables was deduced. A comparison of the experimental (measured from X-ray images) and the predicted (calculated by bone remodelling equation) bone growth data showed good correlation. Overall, the results of this study indicated that significant bone remodelling occurred in the patient with osseointegrated trans-femoral implants. Further, this was could be described well by the bone remodelling equation. Optimisation of the femur modelling method and osseointegration patient gait measurement should contribute more accurate femur models. If such future works can be achieved, further work should be directed towards enhancing the bone remodelling equation and implementing it within an iterative FE framework.

Graphical models reveal the conditional dependence structure between random variables. By estimating the joint density or conditional density, we can detect edges and recover the structure of a graphical model. We propose new nonparametric methods to learn edges for graphical models under a consolidated framework of smoothing spline ANOVA (SS ANOVA) decomposition. We first develop an automatic nonparametric edge detection method by estimating the joint density function with an L1 penalty to interactions in the SS ANOVA decomposition. In the second project, we work directly on the conditional dependence structure and develop a fully nonparametric neighborhood selection method. We detect edges by applying an L1 regularization to interactions in the SS ANOVA decomposition of conditional density functions. These two methods are flexible and contain many existing models as special cases. They also provide a unified framework without any restrictions on the type of each random variable. The joint density approach requires a large computer memory and is thus computationally feasible only when the dimension is small. The neighborhood selection approach overcomes this disadvantage and is more computationally efficient.We propose iterative procedures to compute the estimates and establish the convergence rates for both the joint and conditional density as well as interactions. Simulations indicate that both joint and neighborhood selection methods perform well under Gaussian and non-Gaussian settings. We illustrate the proposed methods using real data examples.

As the per capita disposable income increases, the demand for meat is boosting in China. Pork, accounting for the largest part of Chinese meat consumption, plays an important role in meat demand in China. This study use data collected from a 2014 survey on Chinese meat consumption; four widely different Chinese cities, Beijing, Changsha, Tianjin and Wuhan are selected to provide a more complete view of Chinese pork and meat demand. We use both OLS model and Tobit model to analyze the effects of consumers' socio-demographic characteristics such as age, gender, income, and education on pork and meat demand. Results show that income, marital status, household size have statistically significant effects on pork and meat demand in China. Household heads with children under 18 years old tend to prefer uncooked meat while households with household heads who identify themselves as urbanites tend to spend a significant larger amount of money on cooked or semi-cooked meat.

This dissertation provides a comprehensive methodology for accurate analysis and design of miniaturized radio frequency (RF) surface acoustic wave (SAW) duplexer package. Full-wave analysis based on the three dimensional (3-D) finite element method (FEM) is successfully applied to model the package. The die model is obtained by combining the acoustics and die busbars parasitics models. The acoustics model is obtained using the coupling-of-models (COM) technique. The die busbars, bonding wires and printed circuit board (PCB) are modeled using full-wave analysis. After that, the models of package, die, and bonding wires are assembled together to get the total response. To take into account the mutual couplings, the methodology is extended to model the package, die busbars, and bonding wires together. The advantages and disadvantages of the methodology are also discussed. Based on the methodology, the Korea personal communication system (KPCS) duplexer is analyzed and designed. The isolation of KPCS duplexer package is significantly improved by redesigning inner ground plane, bonding wire scheme and ground via. A KPCS duplexer package is designed and excellent transmitter to receiver isolation in the transmission band is achieved. Simulation and measurement results are compared, and excellent agreement is found. Although we focus on investigating the methods to improve the isolation, the passband performance is also improved. The methodology is also successfully used for flip chip duplexer. The simulation results from our assembling method match the measurement results very well. Optimization method is applied to improve the transmit band isolation. With the new package and die design, the transmit band isolation can be improved from -53.6 dB to -65.2 dB. Based on the new package, the effect of the Rx ground trace on the isolation is investigated and the transmit band isolation can achieve -67.3 dB with the modification of the Rx ground trace. The technique developed in this dissertation reduces the design cycle time greatly and can be applied to various RF SAW device packages.

Urethral obstruction by silica calculi is a major cause of death of cattle and sheep grazing western ranges of the United States. Similar calculi problems have been found in several other countries including Canada, U.S.S.R. and Australia. Range grasses having silica contents of 4 to 7% (dry basis) appear to be the silica source. In addition, silica urinary calculi have been reported to occur in dogs from silicon of unknown origin, humans consuming large quantities of silicate antacids and rats and guinea pigs experimentally fed tetraethylorthosilicate (TES). Emerick and coworkers have found that supplemental dietary phosphate and acid-forming salts provide rats some protection from silica urolithiasis. This protective effect was beyond those effects attributable to increases in water intake and urine volume. The objectives of my thesis research were (1) to study the calculi protective effect of dietary phosphorus using an animal model consisting of laboratory rats fed diets containing 2% tetraethylort 'hosilicate (TES) [(C2 H 5 o)4 Si], (2) to study in vitro the effect of phosphorus on precipitation of a polysilicic acid-protein complex which is believed to be important to the formation of silica calculi, and (3) to develop a ruminant model for future silica urinary calculi studies.