34254 research outputs found
Sort by
Scalable Accelerator Design and Implementation for Efficient Computation of MobileNet Deep Neural Network Models
No full textDeep Neural Networks (DNN) have been widely used in many Artificial Intelligence (AI) applications such as object detection, and image classification. Low-cost and high-speed DNN hardware accelerators are important to accelerate the convolution operations in DNN models for resource-limited edge devices and embedded systems. Due to the large amount of parameters and computations in DNN models, light-weight models with low complexity of computation and parameters have been proposed recently. For example, MobileNets have much smaller amount of parameters and computations compared with some representative DNN models, such as AlexNet, VGG, and ResNet. Depthwise Separable Convolution (DSC), composed of Depthwise Convolution (DWC) and Pointwise Convolution (PWC), is one of the major convolution types in the light-weight DNN models. But, most recent DNN hardware accelerators targeting for normal convolution (CONV) are not efficient in computing DWC. Thus, this thesis aims to design low-power and high-performance DNN hardware accelerators for efficient computation of different types of convolution operations. The first major work of the thesis is the design of hardware accelerators that can compute more efficiently different types of convolutions, including CONV, DWC and PWC. We will exploit run-time reconfigurable dataflow with different parallelism types and degrees to increase hardware utilization rate for different convolution types and feature map sizes. The second major work is to design scalable hardware accelerators that can be easily generated to meet the design trade-off between speed requirement and hardware resource. The third major work is the design of hardware-friendly arithmetic units to compute various activation functions in MobileNets without too much compromise of classification accuracy
A Study on Retailer's Inventory Prediction by Using Machine Learning
No full textIn recent years, the swift changes in modern consumer demands and the intensifying competition within industries have driven businesses to consistently innovate and launch new products to maintain their competitive edge and fulfill customer needs. However, with the diversification and dramatic increase in product sales, the prediction of inventory quantities becomes increasingly complex, leading to potential difficulties in managing stock levels. Having excess inventory can lead to wastage of resources and capital, while insufficient stock can result in customer dissatisfaction and stress for frontline sales staff, thereby affecting performance. Most historical literature has employed machine learning to explore product quantity prediction to improve inventory management. This research, in addition to focusing on product quantity, also investigates other multiple factors influencing inventory variation and identifies the correlation between these factors and case inventory quantities.
Based on the existing real-case data of inventory quantities, order quantities, sales quantities, procurement quantities, as well as off-peak and peak season data, this study analyzes the relationships between these factors and inventory quantities. We utilize machine learning techniques to investigate these factors' impact on inventory quantities. By examining inventory, order, sales, procurement quantities, and off-peak and peak seasons, we identify interrelationships, trends, and validate their accuracy
Mechanical Properties of Quinary Titanium-rich Multi-principal Element Alloy with Low Density and Heterostructure
No full textIn the early stages, multi-principal element alloys were mainly used to adjust different constituent elements or proportions to change the mechanical properties of the alloy. Recently, it was found that the heterostructure would solve the alloy's trade-off between strength and ductility. In addition, the density of the multi-principal element alloy was generally high, which would be unfavorable in the application of the alloy. Therefore, the primary purpose of this study is to solve the trade-off effect and ductility of the alloy. This research aims to investigate the changes in microstructure and mechanical properties of titanium-rich Ti60Al8(VCrNb)32 alloys with low density after various rolling and rapid thermal annealing.
After analysis, it was found that Ti60Al8(VCrNb)32 alloy undergoes rolling and rapid annealing treatment at various times, some areas of the alloy undergo recrystallization, which is called partial recrystallization. This structure includes two microstructures: recrystallization and deformation bands, there have little differences in the mechanical properties between them. The deformation bands belong to the hard zone, while recrystallization belongs to the soft zone. Therefore, it is preliminarily concluded that this structure should belong to a heterogeneous structure. After measuring the hardness of the alloys, it can be found that this series of alloys have high hardness values. Through tensile testing, it can be concluded that after thermal mechanical treatments, the alloy exhibits excellent mechanical properties (after rapid thermal annealing for 39 seconds, the strength increased to 1184MPa, while the ductility remained at 25%) due to its heterogeneous structure
Kim Jong Un\ue2s Nuclear Deterrence Strategy(2012-2020)
No full textThe international power structure has undergone dramatic changes in recent years, and Northeast Asia is one of the complex regions in today's international politics. The threat posed by North Korea's development of nuclear weapons has become increasingly severe since Kim Jong-un took over as the country's leader. In fact, the nuclear weapon program is accelerated, and more missile tests are launched in North Korea. The international community has failed to persuade North Korea to give up its nuclear weapons, denuclearize Korean Peninsula, and establish a permanent peace policy through economic sanctions, and unilateral and bilateral talks. Based on the arguments of offensive realism, Kim Jong-un's nuclear weapons strategy, economic reforms, and utilization of diplomatic strategies in response to international sanctions on the foundation of Songun politics will be examined in this thesis
How to Improve Work Efficiency and Attitude of Employees by Performance Management System ? - A Case Study of W Company
No full textLow fertility rate has become a trend in Taiwan due to population change. Small and medium-sized enterprises often encounter talent shortage. How to select, use, cultivate and retain talents is an significant issue that every enterprise must face.
This study analyzed human resource management system with case interview based on W Company and discovered the weaknesses through the questionnaire from the employees. The employees also provide suggestions to assist the manager in W Company to develop a complete management system on salary and performance. After discussion in this study, the key points can be concluded as below:
1. Individual performance needs to be connected to the department and the company.
2. The criteria of performance evaluation should be sourced and quantified.
3. Performance interview is a key element of successful performance management.
4. Functional Analysis should be connected to on-the-job training.
5. Performance management should be directly connected to human resources management
The Impact of Interest Rates, Ownership Structure (State-Owned and Private), and Diversification on Banks' Performance: Evidence from Taiwan
No full textThis study examines the impact of changes in the interest rate on bank performance by collecting a sample of 30 banks in Taiwan during the period from 2009 to 2022. The empirical results show that the increase in interest rates significantly improves bank profitability. However, increasing interest rates leads to a decrease in capital adequacy ratio and liquidity. Furthermore, when dividing the overall banks into state-owned banks and private banks, the empirical results reveal that private banks outperform state-owned banks in all CAMEL indicators, indicating that private banks can pursue high profits while maintaining good risk control and liquidity. Finally, the study examines the impact of bank diversification on its performance, and the empirical results suggest that increasing the ratio of net non-interest revenue to total net revenue can significantly improve the bank's return on assets, and has a positive relationship with capital adequacy ratio and liquidity
Photochemically Triggered Surface Reactions of Adsorbed Styryl Azide on Cu(100)
No full textNitrenes are the nitrogen counterpart of carbenes. They are fairly reactive substances, and can only be studied after trapping them in low-temperature matrices. Photolysis and thermolysis of organic azides (RN3) lead to the dissociation of N-N2 bond with formation of molecular nitrogen and nitrenes. Here, styryl azide was dosed onto an Cu(100) surface at 160 K, and the desorption of molecularly adsorbed azide occurred below 250 K. In contrast, when exposed to 365 nm UV light irradiation adsorbed styryl azide decomposed, as evidenced by the disappearance of azido N3 stretching vibration band at 2137 cm-1. Following N2 split-off by photolysis, a collection of possible photo-products might be trapped on the surface, including styrylnitrene, 3-phenyl-1-azirine, 3-phenylketenimine, phenylacetonitrile, and indole. Reflection absorption infrared spectra were simulated with density functional theory calculations regarding these intermediate species on the Cu(100) surface for the comparison to the corresponding experimental spectra. Possible photochemical reaction pathways are deduced
Image style transfer using convolutional neural networks with compressed data
No full textIn recent years, style transfer has become increasingly complex in the field of image processing. The original content images often suffer from blurriness during the style transfer process, and the large data size of color images poses challenges in terms of computation and storage ge. Therefore, this research aims to achieve efficient and accurate style transfer by leveraging compression techniques while improving the quality of the content image to enhance the synthesized images\ue2 results.
This research employs a compression scheme that combines Higher-Order Singular Value ecomposition (HOSVD) and Higher-Order Nonnegative Matrix Factorization (HONMF), along with an extended Singular Value Decomposition (SVD) method, to address the issues f long computation time and high storage requirements in style transfer.
By applying the extended HOSVD and HONMF methods for compressing color images, computational efficiency is improved, and storage space is saved. Additionally, VGG19 model is utilized to extract feature maps for computing style loss and content loss in style transfer, with the style loss incorporating the Gram matrix. By integrating data compression echniques, appropriate compression schemes, and well-designed loss functions, the proposed method achieves style transfer while preserving the characteristics of the original content mages
The mechanism of defects formation in silicon carbide
No full textIn this experiment, the Physical Vapor Transport methodwas used to growth silicon carbide single crystals. During the experiment, the thermal field was maintained at an appropriate temperature gradient and argon (Ar) was used as the carrier gas to make the silicon carbide powder transport up to the seed for growth. Non-destructive defect analysis of crystals after growth, including the use of strong light, Optical microscope, Raman spectroscope and Polarized light.
Observing the etch pit of the seed crystal after etching, through an optical microscope and a Raman spectrometer, it is known that the size of the micropipe after the seed crystal was etched with KOH at 460\uc2\ub0C for 10 minutes is about 25 to 28 \uce\ubcm.
Raman spectroscope characteristic peak is about 786.4 cm-1, and the micropipe density of the seed crystal is about 0.02-0.07 cm-2, and it is 4H-SiC.
When the silicon carbide vapor is severely reflowed, considerable of island growths and micropipes will appear on the crystal surface. The cracks on the crystal surface are caused by the difference in the crystal structure of different polytypes of silicon carbide. When polytypes of silicon carbide and poor adhesion are caused severe back sublimation, carbon inclusions caused by graphite particles, and excessive growth speed will increase the crystal micropipe density
Efficient Computation of Depthwise Separable Convolution and Hardware Implementation for Light Weight Neural Network Models
No full textWith the flourishing development of artificial intelligence (AI), Deep Neural Networks (DNN) models have been widely used in fields such as image classification, speech recognition, and object detection. Many DNN hardware accelerators aim to achieve low power consumption and low latency in embedded systems. However, they face the following challenges. The first is the storage problem due to large amount of data and limited memory capacity, resulting in a significant amount of data movement and high power consumption. The second problem is the speed issue as the excessive computational workload cannot be processed in real-time, leading to reduced performance. Light-weight neural network models such as MobileNets and EfficientNets have effectively reduced the computational complexity and number of parameters compared to previous benchmark models like AlexNet and VGG. While maintaining almost the same level of accuracy, these light-weight models have greatly improved overall speed. The core computations of MobileNets and EfficientNets are Depthwise Separable Convolution (DSC), composed of Pointwise Convolution (PWC) and Depthwise Convolution (DWC), which can effectively reduce computation and parameter counts. The first focus of this thesis is to design a hardware architecture that supports multiple light-weight DNN models. The proposed design can improve the low processing element (PE) utilization and large on-chip buffer size encountered in existing hardware architecture designs. The second focus of this thesis is to optimize the data flow for Convolution (Conv), PWC, and DWC operations with merged PWC and DWC to reduce the amount of eexternal memory access. The third focus is the replacement of the hardware-costly swish/sigmoid activation functions by hardware-friendly h-swish/h-sigmoid. Accoring to the experiments on seven model architectures of EfficientNet-V1, we found that the the loss in Top-1 and Top-5 accuracy is within 0.35% and 0.18%, respectively