The growing complexity of neural networks makes their deployment on resource-constrained embedded or mobile devices challenging. With millions of weights and biases, modern deep neural networks can be computationally intensive, with large memory, power and computational requirements. In this thesis, we devise and explore three quantization methods (post-training, in-training and combined quantization) that quantize 32-bit floating-point weights and biases to lower bit width fixed-point parameters while also achieving significant pruning, leading to model compression. We use the total accumulated absolute gradient over the training process as the indicator of importance of a parameter to the network. The most important parameters are quantized by the smallest amount. The post-training quantization method sorts and clusters the accumulated gradients of the full parameter set and subsequently assigns a bit width to each cluster. The in-training quantization method sorts and divides the accumulated gradients into two groups after each training epoch. The larger group consisting of the lowest accumulated gradients is quantized. The combined quantization method performs in-training quantization followed by post-training quantization. We assume storage of the quantized parameters using compressed sparse row format for sparse matrix storage. On LeNet-300-100 (MNIST dataset), LeNet-5 (MNIST dataset), AlexNet (CIFAR-10 dataset) and VGG-16 (CIFAR-10 dataset), post-training quantization achieves 7.62x, 10.87x, 6.39x and 12.43x compression, in-training quantization achieves 22.08x, 21.05x, 7.95x and 12.71x compression and combined quantization achieves 57.22x, 50.19x, 13.15x and 13.53x compression, respectively. Our methods quantize at the cost of accuracy, and we present our work in the light of the accuracy-compression trade-off. / Master of Science / Neural networks are being employed in many different real-world applications. By learning the complex relationship between the input data and ground-truth output data during the training process, neural networks can predict outputs on new input data obtained in real time. To do so, a typical deep neural network often needs millions of numerical parameters, stored in memory. In this research, we explore techniques for reducing the storage requirements for neural network parameters. We propose software methods that convert 32-bit neural network parameters to values that can be stored using fewer bits. Our methods also convert a majority of numerical parameters to zero. Using special storage methods that only require storage of non-zero parameters, we gain significant compression benefits. On typical benchmarks like LeNet-300-100 (MNIST dataset), LeNet-5 (MNIST dataset), AlexNet (CIFAR-10 dataset) and VGG-16 (CIFAR-10 dataset), our methods can achieve up to 57.22x, 50.19x, 13.15x and 13.53x compression respectively. Storage benefits are achieved at the cost of classification accuracy, and we present our work in the light of the accuracy-compression trade-off.

Partial denitrification anammox (PdNA) is an emerging wastewater treatment technology with the potential to increase process capacity and save on energy and carbon. PdNA circumvents potential issues with stability of the more familiar mainstream partial nitritation anammox (PNA) process. The PdNA process can be used to effectively remove ammonia, nitrate, and nitrite from mainstream municipal waste streams. To retain slow growing anammox, some sort of retention system is needed with media being a common solution to this problem. PdNA has been successfully implemented in mainstream full-scale systems in sand filters and with moving media. The goal of this study was to assess the denitrifying capabilities, anammox treatment capacity, and effective surface area to volume of two types of fixed media. A nitrifying pilot was set up to assess the effective surface area to volume. To assess the nitrifying and anammox ammonia removal capabilities of the fixed media, a fixed media PdNA system was installed in the second anoxic zone of a full-scale municipal wastewater treatment plant. The fixed media system consisted of three modules of sheets modified to mimic a plug flow system. After accounting for the estimated nitrate removal from mixed liquor, denitrification rates normalized to media surface area were 0.52 +/- 1.9 g/m2-day in the first module, 0.62 +/- 0.91 g/m2-day for the second module, and 0.56 +/- 0.90 g/m2-day for the third module. In ex situ batch testing it was found that maximum ex-situ anammox ammonia removal rates for the / Master of Science / Urban population growth has created a two-pronged problem for wastewater treatment plants. Plants in populated areas are seeing increases in flow along with growing space restrictions that limit new infrastructure construction. Additionally, rising environmental awareness from the public has spurred regulatory agencies to impose tighter limits on the quality of water leaving plants and entering sensitive watersheds. These factors have driven a need for treatment techniques that allow plants to operate better with their existing equipment. Overall, this concept is known as process intensification. One such method that treatment plants are using to intensify wastewater treatment is the addition of plastic media into their existing tanks. This media provides additional surfaces for the microorganisms that biodegrade the pollutants in the wastewater to grow and allows waste to be treated faster in the same area. It also allows slow growing organisms to be retained in the system that would otherwise not have time to grow. Such slow-growing microbes are especially critical for the removal of ammonia, a toxic form of nitrogen that occurs in high concentrations in wastewater. The partial denitrification-anammox process is an intensification process that leverages microbial metabolisms to convert nitrate to nitrite instead of denitrifying the nitrate all the way to nitrogen gas. Plants then let more ammonia pass through the aeration zone, where ammonia is converted to nitrate. The bleed through ammonia and the nitrite generated from partial denitrification is used by microbes called anammox, which denitrify without the addition of carbon. The full denitrification process requires externally added carbon, which is energy intensive to produce and expensive, and aeration requires energy to run the aeration blowers. Bypassing the full denitrification process using PdNA results in two-fold cost and energy savings. The plastic media help slow-growing anammox bacteria attach and grow to achieve this aim. Most of the plants that use plastic media use media that is free floating in the tank. However, certain plants cannot use this floating plastic media because it can either plug up the system, or flow to the end of the treatment tank and have no way to get back to the front. In instances such as these it could be beneficial to use a type of media that is fixed in place. One potential use of fixed media that has never been tried before is with partial denitrification with anammox. This research sets out to evaluate the effectiveness of fixed media with use in a partial denitrification anammox process and compare it to a treatment tank of moving media that is present at the same plant to find out whether it may be a viable option for retrofitting plants that cannot use moving media.

This paper deals with the pricing differences between a mutual co-operative underwriting system and a fixed-premium underwriting system in providing coverage for marine liability. There has been much debate in recent years within the marine liability underwriting industry over which method fosters more competition, and hence, lower premiums for shipowners who are required to carry such coverage in order to operate. This paper will look at the current mutual marine insurance industry (Protection & Indemnity Associations or P&I Clubs) to compare its pricing both before and after the entry into the market of the fixed-premium underwriters, using data from 1985-2000 that encompasses both a major loss cycle and normal cyclical pricing variations. This analysis will hopefully provide information on whether mutual premium levels for the P&I Clubs differed substantially with the entry of the fixed-price competitors. This is important for the individual shipowners belonging to these mutual underwriting associations, because any variation in premium pricing could mean the potential for either great savings or tremendous losses. / Master of Arts

Fixed bed reactors are used in many different chemical processes, and are a very important part of chemical industry. To model fixed beds we must have a good qualitative understanding of heat transfer in them. Fixed bed models have been developed for high tube-to-particle ratio (N) beds. Modeling of low tube-to-particle beds (3 ¡ÃƒÅ“ N ¡ÃƒÅ“ 8), that are used in extremely exo- and endothermic processes in tube-and-shell type reactors, is complicated, due to the presence of wall effects across the entire radius of the bed. Heat transfer is one of the most important aspects. To obtain accurate models of heat transfer we need to study the physical mechanisms involved especially in the wall vicinity using CFD as a non intrusive tool to collect numerical data. An extra heat transfer resistance is always present near the wall. This is caused by three mechanisms which happen in the wall vicinity. The change of porosity which leads to a change of bed conductivity, the damping of mixing due to the lateral displacement of fluid, the presence of a laminar (viscous) sublayer at the wall. Many authors have been working on how to model the extra resistance near the wall. The main previous approach was to introduce a lumped parameter hw (heat transfer coefficient) which idealizes these three contributions to the extra heat resistance to be at the wall. Our approach will be to keep the parameter hw which will now represent only the viscous boundary layer idealized at the wall, and we are going to incorporate velocity and porosity profiles in the energy equation. In this way we will able to get rid of artificial parameters using the true conductivity of the bed, and the real velocity profile. So we need to study separately each contribution of the different physical mechanisms to clearly understand what happens in the wall vicinity. For this CFD will be a very powerful tool. How CFD models flow near the wall must be understood before starting simulations. Two main approaches for wall bounded flows are available and will be studied: either solve all way down to the wall, or bridge numerical values from the core of the bed to the wall using semi-empirical formulas called wall functions. These methods will be studied and compared. Also with CFD it is possible to run simulations without conduction in the bed, and so, study radial fluid displacement only and obtain reduced velocity profiles. Using the meshing it is also possible to get a very accurate porosity profile. These profiles will be combined in a simplified fixed bed model which will be used to predict temperature profiles. These may then be compared to the full CFD energy solution and to experiment to test the model.

Variation in health care, whether it be in terms of the utilisation of resources, observed health outcomes, costs, quality or access to health care is a well recognised and ever present feature of the modern day health care system. Health care variations challenge basic assumptions about the nature of the health care economy and raise questions about efficiency, equity and where best to direct policy instruments in health care markets. Despite the vast literature documenting variation, and the many discussions around ways to reduce variations in health care markets, the field of dental care has received little interest, in comparison to that of general medical care. This thesis will address this gap and will analyse the variation observed in a specific dental care treatment (dental radiographs) within NHS Scotland, with particular emphasis on the contribution of both dentist and patient unobserved heterogeneity. The thesis takes its focus from two strands of the literature; the underlying theoretical aspect draws on the literature concerning the theory of incentives and physician agency, whilst the empirical component makes use of recent advances in micro-econometric methods, documented in the labour economics literature. Although the thesis is predominantly an empirical analysis, the estimation strategy combines ideas from both the theoretical and empirical literature. A matched patient provider dataset from NHS Scotland is used to conduct an analysis of the variation in dental radiographs, in the presence of, and controlling for unobserved dentist and patient heterogeneity. The results indicate that the remuneration structure alone has little or no impact on the treatment decision to provide a radiograph. When a dentist changes from being on a fixed salary contract to being paid on a fee-for-service basis, they are in fact less likely to provide a radiograph. This result changes in the presence of insurance (identified as being when patients are exempt from the patient charge) and indicates that when the self employed dentist can identify the patient as being exempt, they are more likely to provide a radiograph. This result provides some support for the theory that in the presence of insurance, financial incentives do influence the treatment decision. A final result of the study highlights the importance of accounting for unobserved patient and provider heterogeneity, a factor that has had little attention in the healthcare literature. The results suggest that patient variation, as opposed to the variation across dentists, is much more important in explaining total variation. This is a similar result to that found in both the labour and education literatures.

New information has an important role in asset price movement. This paper investigates the role of scheduled domestic news releases on the Australian government bond market. Specifically, it examines the impact of pre-announced macroeconomic news release on bond futures markets and associated market volatility. Furthermore, an EGARCH-in-mean model is used to determine the asymmetric response of the conditional volatility to either news release or unexpected changes of some news content. The results indicate that excess return of bond futures in the research period was leptokurtic (fat-tailed) with time-varying conditional heteroscedasticity. Day of the week volatility was also present but with a declining pace. It’s generally attributed to the release dates of announcements and information flow from offshore markets. Although announcement effects to the bond futures market were significant, they depended on the type of maturity. Finally, results from EGARCH indicate that fundamental lagging indicators such as CPI and GDP are always important in explaining the impact of news release on market volatility, whereas the unemployment rate has a reasonable role in announcement surprises. The data suggest the following conclusion: investors are seriously concerned with news releases on macroeconomic variables they can feasibly forecast because they are always fundamental and provide a partial indication of the future economy. Surprises from news content are also critical to investors because some important variables can only be forecasted with limited accuracy. Therefore, deviation from anticipated outcomes in the actual content also causes significant market movement. / Thesis(M.Comm.)-- School of Commerce, 2007.

<p>Background: The basis for a price valuation of a company is dependent on both its assets net worth and earnings calculations. While the assets are an important part of the valuation of a company, it is important that the valuation is reliable and resource usage can be measured and allocated to each time period. </p><p>Purpose: The purpose of this essay is to illustrate the difficulties when valuating and allocating own produced fixed assets. </p><p>Accomplishment: We have chosen a case study where the aims are to understand and explain the difficulties with valuating and allocating own produced fixed assets. The research is based on nine personal interviews. </p><p>Result: To obtain a reliable value on own produced fixed asset it requires an accurate project record and a well functioning time management system. The systematic allocation should reflect the actual usage period. An incorrect valuation or systematic allocation of usage can imply that the accounting loses its information value.</p>

Background: The basis for a price valuation of a company is dependent on both its assets net worth and earnings calculations. While the assets are an important part of the valuation of a company, it is important that the valuation is reliable and resource usage can be measured and allocated to each time period. Purpose: The purpose of this essay is to illustrate the difficulties when valuating and allocating own produced fixed assets. Accomplishment: We have chosen a case study where the aims are to understand and explain the difficulties with valuating and allocating own produced fixed assets. The research is based on nine personal interviews. Result: To obtain a reliable value on own produced fixed asset it requires an accurate project record and a well functioning time management system. The systematic allocation should reflect the actual usage period. An incorrect valuation or systematic allocation of usage can imply that the accounting loses its information value.

This research focuses on developing new techniques and designs for highly con- trollable microactuating systems with large force-stroke outputs. A fixed-fixed mi- crobeam is the actuating element in the introduced techniques. Either buckling of a microbridge by thermal stress, lateral deflection of a microbridge by electro- magnetic force, or combined effects of both can be employed for microactuation. The proposed method here is MicroElectroThermoMagnetic Actuation (METMA), which uses the combined techniques of electrical or electro-thermal driving of a mi- crobridge in the presence of a magnetic field. The electrically controllable magnetic field actuates and controls the electrically or electrothermally driven microstruc- tures. METMA provides control with two electrical inputs, the currents driving the microbridge and the current driving the external magnetic field. This method enables a more controllable actuating system. Different designs of microactuators have been implemented by using MEMS Pro as the design software and MUMPs as the standard MEMS fabrication technology. In these designs, a variety of out-of- plane buckling or displacement of fixed-fixed microbeams have been developed and employed as the actuating elements. This paper also introduces a novel actuating technique for larger displacements that uses a two-layer buckling microbridge actu- ated by METMA. Heat transfer principles are applied to investigate temperature distribution in a microbeam, electrothermal heating, and the resulting thermoelas- tic effects. Furthermore, a method for driving microactuators by applying powerful electrical pulses is proposed. The integrated electromagnetic and electrothermal microactuation technique is also studied. A clamped-clamped microbeam carry- ing electrical current has been modeled and simulated in ANSYS. The simulations include electrothermal, thermoelastic, electromagnetic, and electrothermomagnetic effects. The contributions are highlighted, the results are discussed, the research and design limitations are reported, and future works are proposed.

This research focuses on developing new techniques and designs for highly con- trollable microactuating systems with large force-stroke outputs. A fixed-fixed mi- crobeam is the actuating element in the introduced techniques. Either buckling of a microbridge by thermal stress, lateral deflection of a microbridge by electro- magnetic force, or combined effects of both can be employed for microactuation. The proposed method here is MicroElectroThermoMagnetic Actuation (METMA), which uses the combined techniques of electrical or electro-thermal driving of a mi- crobridge in the presence of a magnetic field. The electrically controllable magnetic field actuates and controls the electrically or electrothermally driven microstruc- tures. METMA provides control with two electrical inputs, the currents driving the microbridge and the current driving the external magnetic field. This method enables a more controllable actuating system. Different designs of microactuators have been implemented by using MEMS Pro as the design software and MUMPs as the standard MEMS fabrication technology. In these designs, a variety of out-of- plane buckling or displacement of fixed-fixed microbeams have been developed and employed as the actuating elements. This paper also introduces a novel actuating technique for larger displacements that uses a two-layer buckling microbridge actu- ated by METMA. Heat transfer principles are applied to investigate temperature distribution in a microbeam, electrothermal heating, and the resulting thermoelas- tic effects. Furthermore, a method for driving microactuators by applying powerful electrical pulses is proposed. The integrated electromagnetic and electrothermal microactuation technique is also studied. A clamped-clamped microbeam carry- ing electrical current has been modeled and simulated in ANSYS. The simulations include electrothermal, thermoelastic, electromagnetic, and electrothermomagnetic effects. The contributions are highlighted, the results are discussed, the research and design limitations are reported, and future works are proposed.