Seminar

9:00 am – 5:00 pm

HW 8-28

Title: A learning framework for face detection and region classification

Abstract:

Detection and analysis of faces is a challenging area in computer vision and has been researched actively for applications such as face verification, face tracking, pose estimation. Although current methods based on deep CNNs have achieved remarkable results for the face detection task, it’ still difficult to obtain facial landmark locations and learn hidden relationship among faces simultaneously. The tasks of face detection and face analysis are generally conducted separately. Inspired by current research results, we try to figure out some hidden relationship among human faces. Given a set of multiplayer images of Chinese, observers normally have no idea about individuals come from southern China or Northern China. In this report, we propose a novel framework based on CNNs for simultaneous face detection and region classification from a given image. The network is designed to detect all the faces in an image and classify them into southern group and norther group simultaneously.

Title: Identification of Critical Components in Metro Systems

Abstract:

A metro system provides a customer-centric service, while the delay of waiting on the platform greatly impacts passengers’ travel experience. This paper proposes a novel framework to identify critical components of a metro system by examining the change of waiting punctuality due to possible disruptions or improvement occurring on one component. In static scenarios, candidates of critical components are determined by assessing the average passenger load and origin-destination pairs over one component. Then in dynamic scenarios, these candidates are further evaluated by the change of waiting punctuality. This proposed model is applied to Shenzhen metro network to identify critical components. Results show that the proposed model can be used to develop a component priority classification for practical planning and train scheduling

Title: Real-time information driven graduation manufacturing system of walking-worker assembly islands with fixed-position

Abstract:

Walking-worker assembly islands with fixed-position are layouts in which workers move from one assembly site (often called an assembly island) to another, and required components and tools are moved to an assembly island, while products normally remain in one location for its entire manufacturing (assembly) period according to the process and production plan. Such configuration is not only suitable for producing large, bulky, heavy or fragile products, but also offers considerable flexibility and competitive operational efficiency for products of modest variety and production volumes. However, the configuration of walking-worker assembly islands with fixed-position typically suffers from limited spaces at site and highly dynamics of material and manpower flows in addition to traditional job shop or flow shop problems. Inspired by HKU Graduation Ceremony, a novel model of real-time information driven Graduation Manufacturing System (GMS) is developed to try to solve these problems, and the current progress and future work of the research will be illustrated and discussed.

Title: Demographic correlates of adherence among patients with type 2 diabetes and hypertension

Abstract:

Background - Adequate adherence to therapeutic strategies is important in controlling chronic diseases. However, little is known about the demographic factors associated with the adherence behaviors.

Objective - Our study aimed to examine the correlations between demographic factors and adherence to medication, treatment, and disease-specific activities among patients with type 2 diabetes and hypertension.

Methods - We involved 148 patients diagnosed with Type 2 diabetes and/or hypertension in this study. A questionnaire survey was conducted to collect their demographic information and responses of adherence to medication, treatment and disease-related activities. Multivariate ordered logistic regression was used to identify the correlations between patients’ demographics and adherence behaviors. 

Results - Higher education level was associated with poor medication adherence (OR=0.23; p=0.01). The treatment adherence was higher if the patients were elder (OR= 3.62; p<0.001), if the patients lived with families (OR=3.09; p=0.02), and if the patients had been diagnosed with hypertension for 10-20 years (OR=0.34; p=0.02). No significant relationships of adherence to activities were found.

Conclusions - Considering patients’ education level, age, living status, and disease duration when prescribing therapeutic strategies can promote their adherence behaviors.

Title: Reconstruction of Financial Networks

Abstract:

In this study, we consider the financial network where each node represents a financial institution and it is connected to other nodes via interbank claims. Through this network, the distress of one financial institution may be propagated to other institutions and even affects the non-financial sectors. The risk of such an event to occur is termed as systemic risk in literature.  Most of studies in systemic risk assumes that the structure of the financial network is known. However, in reality, the data of interbank claims are not publicly available, hence the lack of complete information of the financial network.

This study aims to reconstruct the financial networks from partial information. In literature, the maximum entropy (ME) method is widely use to recover the financial network. However, the networks found by the ME method are often quite dense which does not match the empirical observation of a sparse financial network. We target to develop a method that reconstruct the financial network with proper sparsity.

Title: Zero warehouse in manufacture

Abstract:

Warehouses in traditional manufacture are used for storage of various items such as raw materials, work in process inventories and finished products. The basic operations in warehouse include receiving, storage, order picking, and shipping. Considering with the supply chain risks as well as markets demand changes, manufactures often need to hold some raw materials sourcing from their suppliers with a certain amount. After processing, raw materials become work in process inventories stored in the warehouse. Then, finished products come out from the production line are categorized and put up the shelves in warehouses. Then finished products are extracted from a warehousing system to satisfy orders, then products would be shipped in bulk to distributors,

While the inventory costs are always considered as a percentage of inventory value, which depend on the manufacture field and represent nearly 25% of the inventory value on hand.  Any underperformance in warehousing operations can lead to unsatisfactory service and high operational cost. While, along with the rapid growth of production scales and land prices, the shortage of land resources for constructing warehouses is aggravating. Besides, the process of products stored in the warehouse will not add value to the products, which need to be cut down the time as much as possible to save the store cost. The emerging trend is that retailers are sharing demand information with those upstream in their network, and the burden of managing stocks is shifting from the retailer to its suppliers who are capable to respond quickly to changing demand patterns. Therefore, it is imperative to find suitable approaches for improving the products management in manufacture. This research goal is to minimize the warehouse cost in manufacture. The expect result in is to minimize the warehouse in manufacture to a buffer state with a logistic hub to provide warehouse services.

The painting robotic system with mobile base plays an important role in industrial production, considering its flexibility and efficiency for painting operation. The conventional painting robot owning the features of extreme nonlinearity, uncertainty and redundancy makes it one process of high-cost and time consuming. Therefore, the combination for mobile platform and the flexible robotic arm should be accommodated, and simultaneously one task-based optimization for robotic arm should be involved. In this part, one painting robotic system based on the Husky UGV (Unmanned Ground Vehicle) for painting flat wall is adopted, and to improve the performance, one optimization method for a redundant serial robotic manipulator installed on the Husky is proposed. From the details of optimization for serial robot, the optimization process involves the workspace analysis, the optimization for geometry structure and the selection of optimal algorithm. According to the numerical results via the ADAMS software, the whole mobile system can guarantee the performance of stability during the limit operating conditions. And the optimization for serial robot through the PSO (Particle Swarm Optimization) algorithm verifies that the optimized geometric parameters can achieve the best tradeoff for workspace and global velocity performance, which will improve the overall performance for painting in the future.

Abstract:

The universal rule of forecasting is that:

Forecasting is always wrong

The longer the time horizon, the worst the forecast

In this research I attempt to use recent popular Deep Learning (DL) models into forecasting small capitalization stock prices. Why small cap stocks but not index constituents? It is because we believe the small cap space is less efficient, subject to human manipulation and wide fluctuations. It may also exhibit more non-linear dynamics that may be captured by computerized Deep Learning models.

Although the financial market in general are efficient and homogeneous over the long term, the regional sectors such as the small cap market are less efficient and may contain nonlinear dynamics (example: exponential functions, logarithmic functions, etc.) exploitable for forecasting purposes. More traditional econometric models are generally known to work better at longer term time horizon and may be slow in capturing and modeling the patterns in the shorter time horizon in the regional sectors. In this paper, we introduced the recently popular Deep Learning (DL) models into the forecasting area and proposed a DL based forecasting model for small cap stock market. The proposed model adopts an ensemble approach that averages the forecasts using both linear and nonlinear models. ARMA model is used to capture linear data features while DL model is used to capture the nonlinear data feature. Forecasting algorithm under the proposed methodology have been constructed based on different DL models including Multilayer Perception Network (MLP), Deep Belief Network (DBN) and Long Short Term Memory (LSTM) model. We have used the latest small cap market data to evaluate the performance of the proposed forecasting algorithm based on different DL model. We found the ensemble model incorporating DL and ARMA model can somewhat improve the forecasting accuracy. Among different DL models, DBN performs better and can improve the forecast.

Title: Physicians and Nurses’ Perspectives on Performance Feedback Derived from Electronic Health Record Sharing System (eHRSS)

Abstract:

The literature about health information technology mainly focuses on design and implementation of a technology rather than how healthcare staff find the technology useful or easy to use. The fit between healthcare staff and proposed IT is depend on users to accept, reject, use and misuse of a technology.  If healthcare staff find proposed IT useful or easy to use, then an electronic health record would provide great chance to improve quality and safety in healthcare such as reducing errors (medical incidents), saving resources (less repeated test), providing secure and timely access to comprehensive patient records. 

Title: Continuum Manipulators

Abstract:

Apart from revolutionizing the manufacturing sector, robots now have found their way out of the factory and into such applications as surgical manipulators. Minimally invasive surgery (MIS) becomes possible that have less operative trauma, other complications and adverse effects than an equivalent open surgery. So far, the most widespread surgical robot system for minimally invasive surgery is the Da Vinci robot system, a tele-operated robot system, where the physician sits in front of the screen to operate the console and no longer has direct contact with the patient. So the system is designed merely to replicate seamlessly the movement of the surgeon's hands with the functional tips of micro-instruments, not to make decisions automatically. Hence, the trend toward minimally invasive surgery system would be small size, flexible yet strong, and can reach difficult-to-access surgical sites via nonlinear pathways.

Inspired by snake body, robotic researchers now have proposed a kind of slender hyper-redundant manipulator: continuum robots. Continuum robots have enjoyed increasing interest in interventional medicine for their advantages of working in constrained and torturous human lumen. However, due to the inherent nonlinearities and model uncertainties, motion control of such manipulators becomes extremely challenging. Three prevailing designs are tendon-driven continuum manipulators (TCM), tendon-driven serpentine manipulators (TSM) and concentric tube manipulators (CTM). In this seminar, I will compare the three designs at the mechanism level from the kinematics point of view. I introduce relevant research in design, modeling, control, and sensing for continuum manipulators.

Stability, rapidness while walking are most important dynamics aspects in humanoid robots. Co-operative control between lower body and upper torso and arms is expected to improve the dynamics performance mentioned above.

Although some related strategies have been proposed, they can hardly operate in real-time task. We propose the concise but effective algorithm to use upper torso and arm to enhance both stability and rapidness. Based on inverted pendulum model of humanoid robot, an optimal preview control framework of desired ZMP (Zero Moment Point) and actual ZMP is built to ensure the stability of humanoid robot. Also with the compensation of upper body, harmful torque and deviation of the center of mass can be reduced. As a result, the robot can take steps on the stairs and walk with improved stability, rapidness.

Title: Multi-modal Human Haptic Sensing System

Abstract:

We are developing two optic-based multi-modal haptic sensing systems, namely skin sensor and balance sensor. These two sensors operate on an optomechanical sensing principal, that is, Frustrated Total Internal Reflection (FTIR). These two sensors both consist of a high-resolution camera and transparent glass optic-waveguide with either flat or plano-convex shape as well as a light source. The light from the light source is injected into the glass optic-waveguide from its coarse lateral side. Due to the larger refractive index of glass than air, most of light will be reflected back into the optic-waveguide if nothing touches the glass surface. When the sensor touches a sample, diffused reflection happens at all touched points instead of total internal reflection, which will form an optic image in the camera.

The skin sensor is used to examine various of skin conditions. When the sensor is pressed on skin surface gently, the generated haptic image can be used to analyze skin elasticity. Moreover, the 3D skin surface structure is able to be recovered from haptic image, which can be further used to analyze skin roughness. When the sensor is pressed on skin tightly, optic image can be captured to analyze other skin properties like pore size and skin color.

The balance sensor is used for medical diagnosis. When testee stand on the sensor, the camera can capture the variation of force distribution under feet with high resolution and frame rate. Based on the variation information, human balancing capability can be evaluated. The results will be used to diagnose potential neuro diseases. What’s more, we may even be able to figure out the category of disease based on the variation pattern with some deep learning algorithm.

Title: To Automate or Not to Automate, This Is the Question for Ecommerce Logistics: A Game Theoretic Study

Abstract:

The research studies the ecommerce warehousing automation problems in Hong Kong, considering the lease time is too short to justify the investment of small-medium-sized (SME) warehouse operators. There are mainly three stakeholders involved, which are the warehouse property owner (WP), the automation technology provider (WA), and the warehouse operator (WO). In this research, game theory is adopted to model the cooperation and competition relationship among these stakeholders. In the new model, WP will invest in the property development and ask WA to represent itself for collecting rental fee. WA will be responsible for automating the warehouse and recover its investment by sharing WO’s revenue. WO will rent the automated warehouses and the warehouse user, i.e. the ecommerce businesses will pay WO to subscribe automation services. The research has obtained several findings. Firstly, a new business model has been established. By following the suggested cooperation model and pricing mechanism, these stakeholders can gain more profits compared with the current business model. Secondly, the impact of key factors has been discussed through sensitivity analysis. It is found these stakeholders can generate a higher profit given higher turnover rates of items, lower price elasticity and stronger negotiation power. WP in Hong Kong is suggested to sign contracts of shorter lease period for a higher profit in most cases. The research should be of some help in automating the warehousing service industry of Hong Kong by offering some managerial insights. 

Title: Comparison between Markowitz model and Black-Litterman model

Abstract:

Ever since the seminal work of Markowitz (1952), the mean-variance (MV) model has been widely used in the financial industry. However, the practitioners often find that the asset allocation suggested by the MV model not well diversified and it is quite sensitive to the estimation error of input parameters. To address these difficulties, Black and Litterman (1990) propose an improved model that combines investor views with market equilibrium return to identify the optimal asset allocation. In practice, the Black-Litterman (BL) model provides the asset allocation which is more diversified and robust against the estimation error than the one found by the MV model. The objective of this study is to provide the theoretical foundation to this empirical observation. In particular, we target to find under which risk averse level of an investor, the BL model obtains the better asset allocation than the MV model in terms of diversification and profit and when the vice-versa holds. Based on these findings, we aim to propose a novel asset allocation model that further improves the performance and practicality of the BL model. 

As the rising concern of global warming among governments and citizens, Green Supply Chain (GrSC) has been required to reduce Green House Gas (GHG) emissions and satisfy the carbon emission policies over the world. This paper proposes a programming model, which combines the outsourcing decisions with Green Supply Chain design, integrated with the situation under different carbon policies. On one hand, the paper focuses on the warehouse outsourcing decisions making for Green Supply Chain management, while the warehouse outsourcing service is newly generated, such as Cainiao in China, there is no research did before. On the other hand, there are two main carbon emission policies carried out around the world, carbon trading mechanism and carbon pricing policy. China started setting test plots in seven cities for implementing carbon trading policy. It is estimated that a mixed environment of implementing two kinds of policies will form in the future China, which has been applied in some countries and areas, such as France. The paper designs the model practical for dynamics of policy uncertainty. The objective is to minimize the enterprises total costs, including its operation costs and the carbon emission costs, to achieve lowest variable on the enterprise’s profits and reduction of the carbon emissions.

By applying the model with real-life data, we do the numerical analysis and give suggestions to enterprises for the warehouse outsourcing decision making. It has meaningful impact on improvement of enterprises’ logistics performance and environmental sustainable development.

In recent years, human exploration of the world has turned from the macro scale into the micro scale and the study of Nanotechnology has received a constant attention. Nanotechnology, a manipulation of matter on atomic, molecular and supramolecular scale, involves systems on the order of one-thousandth the thickness of a human hair. Micro-bubble and Nano-bubble have gained a lot of attention for its unique properties and wide potential in nano-research field. In this presentation, Ms. YU Huiyang will briefly do a literature review about how bubbles are applied in nanotechnology, also introducing her own work on Nano area.