Дисертації з теми "TRANSFER LEARNING APPROACH"

Diabetic retinopathy (DR) is a complication of diabetes and is a disease that affects the eyes. It is one of the leading causes of blindness in the Western world. As the number of people with diabetes grows globally, so does the number of people affected by diabetic retinopathy. This demand requires that better and more effective resources are developed in order to discover the disease in an early stage which is key to preventing that the disease progresses into more serious stages which ultimately could lead to blindness, and streamline further treatment of the disease. However, traditional manual screenings are not enough to meet this demand. This is where the role of computer-aided diagnosis comes in. The purpose of this report is to investigate how a convolutional neural network together with transfer learning can perform when trained for multiclass grading of diabetic retinopathy. In order to do this, a pre-built and pre-trained convolutional neural network from Keras was used and further trained and fine-tuned in Tensorflow on a 5-class DR grading dataset. Twenty training sessions were performed and accuracy, recall and specificity were evaluated in each session. The results show that testing accuracies achieved were in the range of 35% to 48.5%. The average testing recall achieved for class 0, 1, 2, 3 and 4 was 59.7%, 0.0%, 51.0%, 38.7% and 0.8%, respectively. Furthermore, the average testing specificity achieved for class 0, 1, 2, 3 and 4 was 77.8%, 100.0%, 62.4%, 80.2% and 99.7%, respectively. The average recall of 0.0% and average specificity of 100.0% for class 1 (mild DR) were obtained because the CNN model never predicted this class.
Diabetisk näthinnesjukdom (DR) är en komplikation av diabetes och är en sjukdom som påverkar ögonen. Det är en av de största orsakerna till blindhet i västvärlden. Allt eftersom antalet människor med diabetes ökar, ökar även antalet med diabetisk näthinnesjukdom. Detta ställer högre krav på att bättre och effektivare resurser utvecklas för att kunna upptäcka sjukdomen i ett tidigt stadie, vilket är en förutsättning för att förhindra vidareutveckling av sjukdomen som i slutändan kan resultera i blindhet, och att vidare behandling av sjukdomen effektiviseras. Här spelar datorstödd diagnostik en viktig roll. Syftet med denna studie är att undersöka hur ett faltningsnätverk, tillsammans med överföringsinformation, kan prestera när det tränas för multiklass gradering av diabetisk näthinnesjukdom. För att göra detta användes ett färdigbyggt och färdigtränat faltningsnätverk, byggt i Keras, för att fortsättningsvis tränas och finjusteras i Tensorflow på ett 5-klassigt DR dataset. Totalt tjugo träningssessioner genomfördes och noggrannhet, sensitivitet och specificitet utvärderades i varje sådan session. Resultat visar att de uppnådda noggranheterna låg inom intervallet 35% till 48.5%. Den genomsnittliga testsensitiviteten för klass 0, 1, 2, 3 och 4 var 59.7%, 0.0%, 51.0%, 38.7% respektive 0.8%. Vidare uppnåddes en genomsnittlig testspecificitet för klass 1, 2, 3 och 4 på 77.8%, 100.0%, 62.4%, 80.2% respektive 99.7%. Den genomsnittliga sensitiviteten på 0.0% samt den genomsnittliga specificiteten på 100.0% för klass 1 (mild DR) erhölls eftersom CNN modellen aldrig förutsåg denna klass.

Le but de cette thèse est de minimiser la perte de performance d'un système de détection lorsqu'il rencontre un changement de distribution de données à la suite d’un événement connu (maintenance, ajout de capteur etc.). L'idée est d'utiliser l'approche d'apprentissage par transfert pour exploiter l'information apprise avant l’événement pour adapter le détecteur au système modifié. Un modèle d'apprentissage multitâche est proposé pour résoudre ce problème. Il utilise un paramètre pour équilibrer la quantité d'informations apportées par l'ancien système par rapport au nouveau. Ce modèle est formalisé de manière à pouvoir être résolu par un SVM mono-classe classique avec une matrice de noyau spécifique. Pour sélectionner le paramètre de contrôle, une méthode qui calcule les solutions pour toutes les valeurs du paramètre introduit et un critère de sélection de sa valeur optimale sont proposés. Les expériences menées dans le cas de changement de distribution et d’ajout de capteurs montrent que ce modèle permet une transition en douceur de l'ancien système vers le nouveau. De plus, comme le modèle proposé peut être formulé comme un SVM mono-classe classique, des algorithmes d'apprentissage en ligne pour SVM mono-classe sont étudiés dans le but d'obtenir un taux de fausses alarmes stable au cours de la phase de transition. Ils peuvent être appliqués directement à l'apprentissage en ligne du modèle proposé
The aim of this thesis is to minimize the performance loss of a one-class detection system when it encounters a data distribution change. The idea is to use transfer learning approach to transfer learned information from related old task to the new one. According to the practical applications, we divide this transfer learning problem into two parts, one part is the transfer learning in homogenous space and the other part is in heterogeneous space. A multi-task learning model is proposed to solve the above problem; it uses one parameter to balance the amount of information brought by the old task versus the new task. This model is formalized so that it can be solved by classical one-class SVM except with a different kernel matrix. To select the control parameter, a kernel path solution method is proposed. It computes all the solutions along that introduced parameter and criteria are proposed to choose the corresponding optimal solution at given number of new samples. Experiments show that this model can give a smooth transition from the old detection system to the new one whenever it encounters a data distribution change. Moreover, as the proposed model can be solved by classical one-class SVM, online learning algorithms for one-class SVM are studied later in the purpose of getting a constant false alarm rate. It can be applied to the online learning of the proposed model directly

Learning and development (L&D) professionals in a Fortune 500 company were unable to determine whether managers who completed leadership development courses were transferring what they learned to their work practices. The purpose of this qualitative single instrumental case study was to uncover the factors that accelerated or impeded the transfer of training for employees in the workplace. The conceptual framework was social cognitive learning theory with emphasis on the triadic reciprocal causation model. Guiding questions were used to explore 2 areas: (a) how managers described their preparedness to transfer the training to their jobs, and (b) how managers described their perceptions of the transfer of training from the concepts learned in class to practical job application. Data were collected through one-on-one online interviews with 12 managers who had completed a leadership development course. Data analysis included organizing the data; reading them multiple times; developing codes, categories, and themes; and interpreting the findings. Over 90% of the participants stated that they felt prepared to implement the training after the class. However, only half reported a moderate to high level of confidence incorporating the training into their work. A 3-day professional development project was designed to heighten awareness of the benefits of advancing the transference and application of training with a strong focus on driving social change in the workplace through improved interpersonal skills between managers and their direct reports.

Laser powder bed fusion (LPBF) remains a predominately open-loop additive manufacturing process with minimal in-situ quality and process control. Some machines feature optical monitoring systems but lack automated analytical capabilities for real-time defect detection. Recent advances in machine learning (ML) and convolutional neural networks (CNN) present compelling solutions to analyze images in real-time and to develop in-situ monitoring. Approximately 30,000 selective laser melting (SLM) build images from 31 previous builds are gathered and labeled as either “okay” or “defect”. Then, 14 open-sourced CNN were trained using transfer learning to classify the SLM build images. These models were evaluated by F1 score and down selected to the top 3 models. The top 3 models were then retrained and evaluated using Dietterich’s 5x2 cross-validation and compared with pairwise student t-tests. The pairwise t-test results show no statistically significant difference in performance between VGG- 19, Xception, and InceptionResNet. All models are strong candidates for future development and refinement. Additional work addresses the entire model development process and establishes a foundation for future work. Collaborations with computer science students has produced an image pre-processing program to enhance as-taken SLM images. Other outcomes include initial work to overlay CAD layer images and preliminary hardware integration plan for the SLM machine. The results from this work have demonstrated the potential of an optical layer-wise image defect detection system when paired with a CNN.

Exploitation of the knowledge generated by university research can bring social and economic benefits; thus, knowledge transfer between universities and industry is an important aspect of public policy. In many countries, including the United Kingdom (UK), universities have been developing the capacity to support the commercialisation of publicly funded research, typically by setting up centralised Knowledge Transfer Offices (KTOs). Previous studies have revealed that KTOs need a wide range of abilities to support the commercialisation of academic research, but our understanding of how these abilities are developed and have evolved over time remains limited. In order to address this identified gap in the literature, this thesis examines the questions: What do KTOs learn? How do KTOs learn? and Why do KTOs learn? To address these questions, the thesis adopts a practice-based view of organisational knowledge and learning. The conceptual framework developed to investigate learning by KTOs assumes that their commercialisation practice is learnt through the interactions of their staff within communities of practice, within networks of practice and across communities of practice, and that this learning can be initiated by KTO staff or by targeted strategies devised by the KTO and the university's management. This conceptual framework guides the case studies of six purposefully selected KTOs in the UK. The selection of KTOs is aimed at identifying cases with different learning patterns in order to maximise insights gained from cross-case comparisons as well as at literal replication of the findings. The analysis is based on data collected from semi-structured interviews with key staff in selected KTOs and on information from relevant documents, and follows the ‘explanation building' technique (Yin, 2009). The findings reveal that KTOs tend to develop one of two types of commercialisation practice – each of which is based on different implicit assumptions about generating science-based innovation, and associated with a different set of abilities. Moreover, the findings demonstrate the processes by which changes in practice come about, highlighting the interplay between situated learning and strategic practices of management. The results presented address the aforementioned gap in the literature on university-industry knowledge transfer and contribute to the developing situated learning theory by shedding light on how incremental and more radical changes in practice emerge. The findings should be useful to policy-makers who seek to support universities to build capability for knowledge transfer.

Space debris is becoming an increasingly prevalent issue through a combination of the recent rise in the accessibility of space and the difficulty in actively removing space debris from orbit. The high relative velocity between orbital objects and the difficulty in maintaining their state, results in space debris posing a significant collision risk to active satellites. Risk mitigation strategies rely on space situational awareness, which focuses on tracking space objects and predicting their future states to then inform satellite operators of potential future conjunctions. However, the accuracy of these predictions is limited by a lack of knowledge about the physical characteristics of space debris. This thesis outlines a data-driven approach to space object characterisation through the application of neural networks to light curves extracted from non-resolved ground based optical observations. A light curve is a temporal history of an object's brightness, which contains information about its physical characteristics. Neural networks are more effective when they are trained on a large well-labelled dataset, enabling the complex non-linear relationships within the data to be learned. This has been a limiting factor when applying deep learning to light curve based object classification as light curves are difficult to obtain and label, so real world datasets remain small. This thesis presents simulation-based transfer learning as a method for overcoming this limitation and improving shape classification performance on real world light curve datasets. To further improve performance on challenging cases, a framework for effectively combining multiple light curve observations of a single object is also developed. Finally, a targeted scheduling process has been developed to utilise this framework efficiently, using uncertainty quantification of the neural network output, to selectively prioritise the re-observation of challenging cases and thus reduce misclassifications.

In emerging economies, Small and Medium-Sized Enterprises (SMEs) are threatened by continuous political and economic changes. In such uncertain environments, knowledge is the distinctive factor for the achievement of a competitive advantage. However, limited funds and pressure from competitors force SMEs to seek for external sources of knowledge.

The Multinational Corporation (MNC) represents an alternative for business sustainability within the value chain, including both suppliers and clients. In the aim for pursuing such endeavor, a conceptual framework including inter-firm knowledge transfer processes from the MNC and experiential learning enhanced by the Academia is explored.

In sum, this dissertation is intended to examine the MNC’s and Academia’s role on the procurement of SMEs’ business sustainability through inter-firm knowledge transfer and experiential learning, in terms of absorptive capacity. More specifically, the impact of technical and technological knowledge transferred from the MNC on one side; and reflective learning on managerial skills and business vision from the Academia on the other side, is analyzed through SMEs’ absorptive capacity. Regarding business sustainability, the effect of the application of newly absorbed knowledge is analyzed in terms of SMEs’ selected indicators for business improvements. As a complement, a qualitative study is included in order to provide support for findings hereby obtained.

The learning using privileged information paradigm has allowed support vector machine models to incorporate privileged information, variables available in the training set but not in the test set, to improve predictive ability. The consequent introduction of the knowledge transfer method has enabled a practical application of support vector machine models utilizing privileged information. This thesis describes a modified knowledge transfer method inspired by cross-validation, which unlike the current standard knowledge transfer method does not create the knowledge transfer function and the approximated privileged features used in the support vector machines on the same observations. The modified method, the robust knowledge transfer, is described and evaluated versus the standard knowledge transfer method and is shown to be able to improve the predictive performance of the support vector machines for both binary classification and regression.

In recent years, there has been an increase in women securing leadership positions across Canadian sport. However, when compared with their male counterparts, there continues to be an imbalance of women in these roles. The purpose of this doctoral dissertation was to evaluate a social learning initiative implemented in the province of Alberta to address these existing gender disparities by increasing gender equity, leadership development/diversity, and knowledge transfer across sport systems. The Alberta Women in Sport Leadership Impact Program (AWiSL) was framed using Wenger’s (1998) concept Communities of Practice and consisted of 12 sport leaders (from various PSOs, clubs, and other sport organizations) and six mentors (with leadership expertise). Each sport leader planned and implemented a project in their home sport organizations to support the increase of gender equity and leadership development/diversity. The mentors were responsible for supporting the sport leaders in achieving their project goals and facilitating leadership development opportunities to inspire growth in the sport leaders. Accordingly, an evaluation was conducted using the Value Creation Framework (Wenger-Trayner et al., 2011) to examine the perceived value of participating in this social learning initiative. Data were collected over a year and a half period, from the 18 members who made up the AWiSL group and other important stakeholders. The data included in-depth interviews, informal conversations, observations, surveys, and collecting organizational documents resulting in over 700 pages of transcribed data. The findings are presented in four articles and an additional findings section. The first article focuses on one of the sport leader’s projects which aimed to foster a collaborative women-only training program for 10 women to become certified coach developers. The second article examines the development of the AWiSL mentors’ social learning leadership capabilities during their first attempt at facilitating a CoP to promote gender equity and leadership development/diversity, through an action learning approach. The third article delves into the sport leaders’ perceptions of their leadership skill development through their participation in the two and a half year social learning initiative, specifically a CoP of femininity. Finally, the fourth article highlights the 12 sport leaders’ projects to examine the impacts of the AWiSL in terms of moving gender equity forward across the province. The additional findings section touches on the knowledge transfer outcome of the AWiSL, including the development of a how-to model for organizations wishing to implement a similar initiative and the overall perceived value of this initiative. The dissertation is concluded with a general discussion highlighting the theoretical contributions and practical implications, along with future recommendations for research.

This thesis is concerned with receptivity and response encountered at different levels within organisations when a novel approach to the learning of fault diagnosis skills is introduced. Essentially, the work involved the transfer of a learning technology from research and development on the one hand to the workplace on the other. With only a few exceptions, previous research had taken a highly focused, machinecentred view of fault diagnosis. The same view has been adopted towards the limited range of training that is currently offered in this subject. The overall aim here was to introduce a holistic approach by viewing fault diagnosis as a social process that is conducted within a technical context. To do this, account had to be taken of the complex interactions found between a number of disciplines such as, design, production, quality assurance, buying, maintenance and management. The learning technology that served as a vehicle for the transfer of this systems approach was a series of open learning modules. The modules were produced as part of the project. The methodology was based upon an inductive approach that involved the interpretation of qualitative data; this was done using a triangulation of research methods: case studies, critical incidents, and survey questionnaire. The sample, of both large and small organisations, was designed to provide a mix of different types of manufacturing and service industries. In each case, the practice of fault diagnosis skills continues to be a critical influence upon business performance. Different factors arose at different levels within each organisation, and betweenorganisation factor differences are also identified. Apart from the production of open learning material, the contribution made to the subject area is of new insights into the mechanism used for technology transfer within companies, and the identification of factors that either facilitate or hinder transfer of this kind. There is also a contribution to the debate about how the theory of systems thinking can be applied in a prescriptive way as opposed to the more common descriptive delivery. Recommendations are made for further developmento f the learning technology.

Educational Digital Libraries (DLs) are complex information systems which are designed to support individuals' information needs and information seeking behavior. To have a broad impact on the communities in education and to serve for a long period, DLs need to structure and organize the resources in a way that facilitates the dissemination and the reuse of resources. Such a digital library should meet defined quality dimensions in the 5S (Societies, Scenarios, Spaces, Structures, Streams) framework - including completeness, consistency, efficiency, extensibility, and reliability - to ensure that a good quality DL is built. In this research, we addressed both external and internal quality aspects of DLs. For internal qualities, we focused on completeness and consistency of the collection, catalog, and repository. We developed an application pipeline to acquire user-generated computing-related resources from YouTube and SlideShare for an educational DL. We applied machine learning techniques to transfer what we learned from the ACM Digital Library dataset. We built classifiers to catalog resources according to the ACM Computing Classification System from the two new domains that were evaluated using Amazon Mechanical Turk. For external qualities, we focused on efficiency, scalability, and reliability in DL services. We proposed cloud-based designs and applications to ensure and improve these qualities in DL services using cloud computing. The experimental results show that our proposed methods are promising for enhancing and enriching an educational digital library. This work received support from ACM, as well as the National Science Foundation under Grant Numbers DUE-0836940, DUE-0937863, and DUE-0840719, and IMLS LG-71-16-0037-16.
Ph. D.

Dans l'état actuel des connaissances, il est impossible de poser correctement toute la physique permettant de modéliser les hydrosystèmes dans leur ensemble, notamment à cause de la dynamique très contrastée des différents compartiments. Les modèles systémiques simplifient la représentation des hydrosystèmes en ne considérant que leurs flux d’échange. L’objet de ce travail est de proposer un outil de modélisation systémique fournissant des informations sur le fonctionnement physique des hydrosystèmes, tout en étant simple et parcimonieux. Ce modèle nommé MASH (pour Modélisation des Hydrosystèmes par Approche Systémique) est basé sur l’utilisation de fonctions de transfert paramétriques choisies en fonction de leur faible paramétrisation, leur caractère général et leur interprétation physique. Il est versatile, dans le sens que son architecture est modulable et que le nombre d’entrées, le nombre de fonctions de transfert en série et le type de fonctions de transfert utilisé est laissée à la discrétion de l’utilisateur. Ce modèle est inversé en utilisant de récentes avancées en apprentissage automatique grâce à une famille d’heuristiques basée sur l’intelligence en essaim nommé « optimisation par essaim de particule » (ou PSO pour « Particle Swarm Optimization »). Le modèle et ses algorithmes d’inversion sont testés sur un cas d’école synthétique, puis sur un cas d’application réel
In the light of current knowledge, hydrosystems cannot be modelled as a whole since underlying physical principles are not totally understood. Systemic models simplify hydrosystem representation by considering only water flows. The aim of this work is to provide a systemic modelling tool giving information about hydrosystem physical behavior while being simple and parsimonious. This model, called HMSA (for Hydrosystem Modelling with a Systemic Approach) is based on parametric transfer functions chose for their low parametrization, their general nature and their physical interpretation. It is versatile, since its architecture is modular, and the user can choose the number of inputs, outputs and transfer functions. Inversion is done with recent machine learning heuristic family, based on swarm intelligence called PSO (Particle Swarm Optimization). The model and its inversion algorithms are tested first with a textbook case, and then with a real-world case

Active Approaches provide chances for individual, comprehension-oriented learning and can facilitate the acquirement of general mathematical competencies. Using the example of pinboards, which were developed for different areas of the secondary level, workshop participants experience, discuss and further develop learning tasks, which can be used for free activities, for material based concept formation, for coping with heterogeneity, for intelligent exercises, as tool for the presentation of students’ work and as basis for games. The material also allows some continuous movements and can thus prepare an insightful usage of dynamic geometry programs. Central Part of the workshop is a work-sharing group work with learning tasks for grades 5 to 8. The workshop will close with a discussion of general aspects of material-based learning.

In an Information and Communication Technology (ICT) system, information and knowledge have a key role in the development as well as in the evolution of processes. Due to the continuous improvement of the ICT, there are no limits on when, where and how each process has to take place. This condition, if on one hand permits that all the processes can be easily performed, on the other hand it increases the complexity level of each process itself. Furthermore the development of each process is much more complicated considering the concepts of social networking. In fact, taking into consideration the mechanism of social influence and social contagion as well as the capability and knowledge of each individual, the network node is affected, in a positive or in a negative way, not only by the other nodes of the network connected to it but also by its position and importance within the network. Considering an ICT system, there are a lot of processes that can take place within a network. The main focus of my Ph. D. research activity has been to analyse the decision making process taking place in a social network and, in particular, the main features that influence the development of the process itself. In fact, due to events and objectives that an individual, the decision maker (DM), had to face and deal, it becomes necessary to take decisions. In an ICT system, each decision making process is characterised by four main features: dynamism, context-dependence, multiple criteria and social influence. Dynamism expresses the continuous change of the characteristics of both environment and of the decision maker who has to perform the process at each time step. Context-dependence, instead, means the importance of the context, defined as the information that is necessary to describe the situation where a decision maker performs its processes. As expressed before and as a confirmation of the importance of a multiple criteria decision analysis, the paradigm of the decision making process is the evaluation of each alternative on the basis of a set of criteria. In this way the advantages and the drawbacks of each alternative are highlighted. Social influence has to be taken into consideration in the development of the decision making process, because the decision maker performs its process not alone but it is surrounded by other individuals that have a minor or a greater, a positive or a negative, influence on it, leading its decisions near or far from its initial inclination, as a results of social interactions among individuals. These four aspects have to be considered together with the personal features of each decision maker, like, for example, its psychological and psycophysical state. Thus, considering the aspects previously introduced, this Ph. D. dissertation proposes a multiple criteria and context-aware decision making model being able to represent the decision making process of an individual in a social network. This model is able to represent the dynamics of decision taken by an individual within a social network, considering the variation of the context and the influence that the individual perceives from its neighborhood. The behaviour of each individual is represented by a set of parameters, whose variation influences the dynamics of decision within the social network. Successively, applying the same perspective to the process of knowledge transfer and learning, it is possible to consider these processes as individual decision making process where each individual has to decide if accept or not knowledge from its neighboring nodes. In the Ph. D. dissertation the concepts and the analytical instruments provided by the multiple criteria decision analysis (MCDA) are applied to social networks in order to represent as much as possible realistic decision making processes involving individuals that are parts of social networks in different contexts.

Interest in food and nutrition education is growing in response to the need for a focus on health within the community. It has in the past however been largely overlooked in schools. Schools have the opportunity to influence food decisions by cost effectively reaching a wide audience who have an interest in food. Present education reforms in Australia offer the opportunity for food and nutrition education to take its vital place in the school setting. Food and nutrition promotion in schools is often perceived as curriculum based. This is neither the best nor the only approach. Tuckshops, the wider school community and the media play key roles in food and nutrition education. This research uses the voices of participants in the study to explore the realities and expectations surrounding food and nutrition education in schools today. The study is referred to under the acronym FAST (Food and Schools Today) and draws on data from the background NEAT (Nutrition Education and Teenagers) project. A HPS (Health Promoting Schools) approach is highlighted as the preferred strategy for achieving long term behavioural benefits and heightened awareness of food-related issues. This approach was incorporated into the Nutrition Success cycle and trialed in twelve Queensland schools. Participant ideas and enthusiasm for school based initiatives were encouraging. When individuals inquire about food and nutrition issues autonomously, they are more likely to choose appropriate information to suit their specific needs. Such a broad philosophy involving critical thinking is therefore, the most important ingredient for successful food and nutrition promotion. The strength of groups working collaboratively to trial the Nutrition Success cycle, lay in their ability to actively involve a wide range of autonomous, motivated individuals with clear goals.

Cette recherche porte sur la co-construction d'une stratégie pour implanter dans la pratique infirmière une échelle d'évaluation du risque de constipation du patient hospitalisé (ERCoPH). Elle s'appuie sur le Modèle humaniste des soins infirmiers (Girard et Cara, 2011) et sur le modèle d'apprentissage socio-constructivisme (Vygotsky, 1997). Le design s'inscrit dans une approche collaborative (Desgagné, 1997). Les objectifs sont de co-construire une stratégie pour implanter cette nouvelle échelle et d'évaluer l'impact de cette approche sur le développement professionnel continu (PDC) des infirmières ayant participé à cette étude et sur le raisonnement clinique de leurs pairs. Cette approche a permis à un groupe d'infirmières lors de séances d'analyse en groupe (Van Campenhoudt &al. 2005) de modéliser des perspectives pour implanter l'échelle ERCoPH. L'impact sur le DPC des équipes non participantes s'est appuyé sur une enquête avant-après. A partir de l'observation d'entretiens d'accueil de patients hospitalisés et d'une enquête sur la capacité à catégoriser les patients à risque de constipation. L'approche collaborative a entrainé chez les infirmières du groupe collaboratif un développement professionnel, notamment dans leurs capacités réflexives. La co-construction de cette stratégie d'implantation de l'échelle ERCoPH peut être associé à un modèle de transfert de connaissances tel que défini par Fixsen et al. (2005) et Graham et al. (2006) dont le but est de permettre l'intégration dans la pratique de nouvelles connaissances et ainsi réduire les écarts avec la pratique
This study focuses on the co-construction of a strategy aiming to implement, in nursing practice, a rating scale to assess the risk of constipation in hospitalised patients (ARCoPH). It is based on humanistic model of nursing (Girard et Cara, 2011) and on the social constructivist approach to learning (Vygotsky, 1997). The research design uses a collaborative approach (Desgagné, 1997). The objectives are to co-construct a strategy to implement this new scale and the impact of this approach on the continuing professional development (CPD) of nurses who participated in the study and on the clinical reasoning of their peers. Using a collaborative approach, a group of five nurses developed, during group analysis sessions (Van Campenhoudt et al., 2005), practical insights to implement the ARCoHP scale. The impact on their CPD was determined through a group interview and a questionnaire. The effect of this approach on the clinical reasoning of the teams was established using a before and after survey based on the observation of patient intake interviews, and to assess the nurses' ability to identify patients at risk of constipation. This collaborative approach led to the professional development of participating nurses, specifically to the improvement of their reflective skills.The co-construction of this implementation strategy for the ARCoHP scale can be associated with the transfer of learning model as defined by Fixsen et al. (2005) and Graham et al. (2006), and thus help close the gaps between theory and practice

Une des promesses des plateformes de réseaux sociaux (comme Twitter et Facebook) est de fournir un endroit sûr pour que les utilisateurs puissent partager leurs opinions et des informations. Cependant, l’augmentation des comportements abusifs, comme le harcèlement en ligne ou la présence de discours de haine, est bien réelle. Dans cette thèse, nous nous concentrons sur le discours de haine, l'un des phénomènes les plus préoccupants concernant les réseaux sociaux.Compte tenu de sa forte progression et de ses graves effets négatifs, les institutions, les plateformes de réseaux sociaux et les chercheurs ont tenté de réagir le plus rapidement possible. Les progrès récents des algorithmes de traitement automatique du langage naturel (NLP) et d'apprentissage automatique (ML) peuvent être adaptés pour développer des méthodes automatiques de détection des discours de haine dans ce domaine.Le but de cette thèse est d'étudier le problème du discours de haine et de la détection des propos injurieux dans les réseaux sociaux. Nous proposons différentes approches dans lesquelles nous adaptons des modèles avancés d'apprentissage par transfert (TL) et des techniques de NLP pour détecter automatiquement les discours de haine et les contenus injurieux, de manière monolingue et multilingue.La première contribution concerne uniquement la langue anglaise. Tout d'abord, nous analysons le contenu textuel généré par les utilisateurs en introduisant un nouveau cadre capable de catégoriser le contenu en termes de similarité basée sur différentes caractéristiques. En outre, en utilisant l'API Perspective de Google, nous mesurons et analysons la « toxicité » du contenu. Ensuite, nous proposons une approche TL pour l'identification des discours de haine en utilisant une combinaison du modèle non supervisé pré-entraîné BERT (Bidirectional Encoder Representations from Transformers) et de nouvelles stratégies supervisées de réglage fin. Enfin, nous étudions l'effet du biais involontaire dans notre modèle pré-entraîné BERT et proposons un nouveau mécanisme de généralisation dans les données d'entraînement en repondérant les échantillons puis en changeant les stratégies de réglage fin en termes de fonction de perte pour atténuer le biais racial propagé par le modèle. Pour évaluer les modèles proposés, nous utilisons deux datasets publics provenant de Twitter.Dans la deuxième contribution, nous considérons un cadre multilingue où nous nous concentrons sur les langues à faibles ressources dans lesquelles il n'y a pas ou peu de données annotées disponibles. Tout d'abord, nous présentons le premier corpus de langage injurieux en persan, composé de 6 000 messages de micro-blogs provenant de Twitter, afin d'étudier la détection du langage injurieux. Après avoir annoté le corpus, nous réalisons étudions les performances des modèles de langages pré-entraînés monolingues et multilingues basés sur des transformeurs (par exemple, ParsBERT, mBERT, XLM-R) dans la tâche en aval. De plus, nous proposons un modèle d'ensemble pour améliorer la performance de notre modèle. Enfin, nous étendons notre étude à un problème d'apprentissage multilingue de type " few-shot ", où nous disposons de quelques données annotées dans la langue cible, et nous adaptons une approche basée sur le méta-apprentissage pour traiter l'identification des discours de haine et du langage injurieux dans les langues à faibles ressources
The great promise of social media platforms (e.g., Twitter and Facebook) is to provide a safe place for users to communicate their opinions and share information. However, concerns are growing that they enable abusive behaviors, e.g., threatening or harassing other users, cyberbullying, hate speech, racial and sexual discrimination, as well. In this thesis, we focus on hate speech as one of the most concerning phenomenon in online social media.Given the high progression of online hate speech and its severe negative effects, institutions, social media platforms, and researchers have been trying to react as quickly as possible. The recent advancements in Natural Language Processing (NLP) and Machine Learning (ML) algorithms can be adapted to develop automatic methods for hate speech detection in this area.The aim of this thesis is to investigate the problem of hate speech and offensive language detection in social media, where we define hate speech as any communication criticizing a person or a group based on some characteristics, e.g., gender, sexual orientation, nationality, religion, race. We propose different approaches in which we adapt advanced Transfer Learning (TL) models and NLP techniques to detect hate speech and offensive content automatically, in a monolingual and multilingual fashion.In the first contribution, we only focus on English language. Firstly, we analyze user-generated textual content to gain a brief insight into the type of content by introducing a new framework being able to categorize contents in terms of topical similarity based on different features. Furthermore, using the Perspective API from Google, we measure and analyze the toxicity of the content. Secondly, we propose a TL approach for identification of hate speech by employing a combination of the unsupervised pre-trained model BERT (Bidirectional Encoder Representations from Transformers) and new supervised fine-tuning strategies. Finally, we investigate the effect of unintended bias in our pre-trained BERT based model and propose a new generalization mechanism in training data by reweighting samples and then changing the fine-tuning strategies in terms of the loss function to mitigate the racial bias propagated through the model. To evaluate the proposed models, we use two publicly available datasets from Twitter.In the second contribution, we consider a multilingual setting where we focus on low-resource languages in which there is no or few labeled data available. First, we present the first corpus of Persian offensive language consisting of 6k micro blog posts from Twitter to deal with offensive language detection in Persian as a low-resource language in this domain. After annotating the corpus, we perform extensive experiments to investigate the performance of transformer-based monolingual and multilingual pre-trained language models (e.g., ParsBERT, mBERT, XLM-R) in the downstream task. Furthermore, we propose an ensemble model to boost the performance of our model. Then, we expand our study into a cross-lingual few-shot learning problem, where we have a few labeled data in target language, and adapt a meta-learning based approach to address identification of hate speech and offensive language in low-resource languages

Dosage calculations are an essential skill for current nurse practitioners. The challenge for educators is to determine how the learner can be best assisted to learn and apply both general and specific knowledge and skills. This fundamental issue of providing instruction that is meaningful and able to be utilised in another setting is an age-old concern. It can be seen from the posology literature that outcomes from a diversity of instructional attempts have generally been limited. This exploratory field study investigated the effect of specifically prepared instructional approaches upon learning outcome and also the learners' ability to transfer this knowledge. The research is grounded in both transfer of learning and metacognitive-regulatory theory. Three instructional approaches were developed and considered in this study. The computerised learning approach was designed to encourage the low-road of learning providing for automaticity in skill performance. The integrative learning approach incorporated process-oriented instruction to support high-road learning and also repetitive practice to foster the low-road of learning. In addition, it included small group discussion to address the affective component of mathematical phobia often intrinsic to dosage calculations. The third approach combined the strategies provided in both the computerised and integrative learning approaches. Based upon the literature, it was assumed that the integrative approach would be most effective in developing all forms of knowledge, particularly conditional knowledge, and consequently greater performance in far transfer tasks would be evidenced. Furthermore, this effect would be greatest in those learners who reported a negative attitude towards mathematics and mathematical testing and who also lacked self-regulation or external-regulation of learning. The combination of the computerised and integrative approach was expected to enhance the low-road of learning and consequently greater performance on near transfer tasks would be evidenced. Evaluation used a methodological mix of both quantitative and qualitative approaches. The findings were not entirely conclusive, although they did offer some support to the study claims and interesting insight into other issues that need to be accounted for in exploratory field studies of this type. Overall, it appeared that computerised learning might have been more influential in the development of procedural knowledge. However, when learners reported higher levels of negative attitudes towards mathematics and mathematical testing the integrative approach was more effective than the computerised approach in developing procedural knowledge. There was some evidence to suggest that when the learner reported being highly self-regulated or reliant on external-regulation, procedural knowledge development was interfered with when they received the combination of computerised and integrative learning. Although not statistically proven the integrative approach did result in higher scores on conditional knowledge measures in both the first and second post-tests. However when the effect of negative attitudes towards mathematics and mathematical testing was accounted for, then statistical support was evident, indicating that the integrative approach was more effective than the computerised approach under these circumstances. This effect was also noted when the learner reported a medium level of self-regulation. Generally the reported level of metacognitive-regulation did not appear to influence the treatment effects. None of the treatments examined demonstrated greater effectiveness on measures of far transfer. These results support the findings of earlier studies both within the posology and transfer of learning literature asserting that the phenomenon of transfer can be elusive and does not naturally ensue from attempts made to improve upon instructional approaches. However due to institutional constraints the intervention period in the present study was markedly short affecting the integrity of the conceptual framework underlying the study. Despite this, the statistical evidence from the study suggests that both the integrative and computerised learning approach are worthy inclusions into future instructional approaches aimed towards developing competency in dosage calculations.

Dissertação de mestrado Erasmus Mundus para obtenção do grau de mestre em Técnicas Laboratoriais Forenses
Soil analysis is an important part of forensic science as it can provide vital links between a suspect and a crime scene based on its characteristics. The use of soil in a forensic context can be characterised into two categories: intelligence purposes or court purposes. The core basis of the comparison of sites to determine the provenance is that soil composition, type etc. vary from one place to another. The aim of this project is to ‘map’ soils and predict the location of a sample of unknown origin based on the chemometric profiles of Fourier transform infrared (FTIR) spectra, micro x-ray fluorescence profiles and visible spectra. Thirty one samples were collected in triplicate from Monsanto Park in Lisbon for each predetermined collection point on a defined grid. Full FTIR spectra (400-4000cm-1), Visible (1100-401cm-1) spectra, UV (400-200cm-1) spectra and μXRF profiles were collected for all samples. A subset of 43 discriminant features was selected from a total of 1430 using the Boruta feature selection algorithm from the FTIR, μXRF and visible spectra. These discriminant features acted as input data that was used to create a neural network which allowed the prediction of Cartesian co-ordinates (or location) of the samples with a high degree of accuracy (86%) and has shown to be a very useful approach to predict soil location.

Nell'era dell'informatizzazione e con l'avvento del web semantico vi è una sempre crescente necessità di classificare e organizzare grandi moli di dati non strutturati posti in linguaggio naturale al fine di poter trarre da essi informazioni utili alla costruzione di una conoscenza. Tuttavia l'analisi e la classificazione di tali tipi di dati rappresenta, per un sistema automatico, un problema non triviale quanto per un essere umano. A tale scopo in tempi odierni assistiamo a una larga diffusione di tecniche di natural language processing, text mining e sentiment analysis grazie anche ai recenti importanti sviluppi nel campo del machine learning. L'obiettivo che questo elaborato si pone è quello di presentare un approccio alternativo alla classificazione automatica di documenti in categorie gerarchiche di argomenti basato sulle relazioni semantiche che intercorrono tra le parole che compongono gli stessi. Tali relazioni vengono attinte da una base di conoscenza semantica mediante algoritmi sviluppati appositamente, rendendo così la determinazione del grado di correlazione completamente indipendente dagli argomenti usati per l'addestramento e persino dalla lingua. Il modello proposto si differenzia da quelli presenti in letteratura classica per le sue caratteristiche di riusabilità e generalizzazione grazie alle quali è in grado di operare, in fase di classificazione, anche su tematiche e categorie non conosciute in fase di addestramento.

In the last 50 years, flooding has figured as the most frequent and widespread natural disaster globally. Extreme precipitation events stemming from climate change could alter the hydro-geological regime resulting in increased flood risk. Near real-time precipitation monitoring at local scale is essential for flood risk mitigation in urban and suburban areas, due to their high vulnerability. Presently, most of the rainfall data is obtained from ground‐based measurements or remote sensing that provide limited information in terms of temporal or spatial resolution. Other problems may be due to the high costs. Furthermore, rain gauges are unevenly spread and usually placed away from urban centers. In this context, a big potential is represented by the use of innovative techniques to develop low-cost monitoring systems. Despite the diversity of purposes, methods and epistemological fields, the literature on the visual effects of the rain supports the idea of camera-based rain sensors but tends to be device-specific. The present thesis aims to investigate the use of easily available photographing devices as rain detectors-gauges to develop a dense network of low-cost rainfall sensors to support the traditional methods with an expeditious solution embeddable into smart devices. As opposed to existing works, the study focuses on maximizing the number of image sources (like smartphones, general-purpose surveillance cameras, dashboard cameras, webcams, digital cameras, etc.). This encompasses cases where it is not possible to adjust the camera parameters or obtain shots in timelines or videos. Using a Deep Learning approach, the rainfall characterization can be achieved through the analysis of the perceptual aspects that determine whether and how a photograph represents a rainy condition. The first scenario of interest for the supervised learning was a binary classification; the binary output (presence or absence of rain) allows the detection of the presence of precipitation: the cameras act as rain detectors. Similarly, the second scenario of interest was a multi-class classification; the multi-class output described a range of quasi-instantaneous rainfall intensity: the cameras act as rain estimators. Using Transfer Learning with Convolutional Neural Networks, the developed models were compiled, trained, validated, and tested. The preparation of the classifiers included the preparation of a suitable dataset encompassing unconstrained verisimilar settings: open data, several data owned by National Research Institute for Earth Science and Disaster Prevention - NIED (dashboard cameras in Japan coupled with high precision multi-parameter radar data), and experimental activities conducted in the NIED Large Scale Rainfall Simulator. The outcomes were applied to a real-world scenario, with the experimentation through a pre-existent surveillance camera using 5G connectivity provided by Telecom Italia S.p.A. in the city of Matera (Italy). Analysis unfolded on several levels providing an overview of generic issues relating to the urban flood risk paradigm and specific territorial questions inherent with the case study. These include the context aspects, the important role of rainfall from driving the millennial urban evolution to determining present criticality, and components of a Web prototype for flood risk communication at local scale. The results and the model deployment raise the possibility that low‐cost technologies and local capacities can help to retrieve rainfall information for flood early warning systems based on the identification of a significant meteorological state. The binary model reached accuracy and F1 score values of 85.28% and 0.86 for the test, and 83.35% and 0.82 for the deployment. The multi-class model reached test average accuracy and macro-averaged F1 score values of 77.71% and 0.73 for the 6-way classifier, and 78.05% and 0.81 for the 5-class. The best performances were obtained in heavy rainfall and no-rain conditions, whereas the mispredictions are related to less severe precipitation. The proposed method has limited operational requirements, can be easily and quickly implemented in real use cases, exploiting pre-existent devices with a parsimonious use of economic and computational resources. The classification can be performed on single photographs taken in disparate conditions by commonly used acquisition devices, i.e. by static or moving cameras without adjusted parameters. This approach is especially useful in urban areas where measurement methods such as rain gauges encounter installation difficulties or operational limitations or in contexts where there is no availability of remote sensing data. The system does not suit scenes that are also misleading for human visual perception. The approximations inherent in the output are acknowledged. Additional data may be gathered to address gaps that are apparent and improve the accuracy of the precipitation intensity prediction. Future research might explore the integration with further experiments and crowdsourced data, to promote communication, participation, and dialogue among stakeholders and to increase public awareness, emergency response, and civic engagement through the smart community idea.
Negli ultimi 50 anni, le alluvioni si sono confermate come il disastro naturale più frequente e diffuso a livello globale. Tra gli impatti degli eventi meteorologici estremi, conseguenti ai cambiamenti climatici, rientrano le alterazioni del regime idrogeologico con conseguente incremento del rischio alluvionale. Il monitoraggio delle precipitazioni in tempo quasi reale su scala locale è essenziale per la mitigazione del rischio di alluvione in ambito urbano e periurbano, aree connotate da un'elevata vulnerabilità. Attualmente, la maggior parte dei dati sulle precipitazioni è ottenuta da misurazioni a terra o telerilevamento che forniscono informazioni limitate in termini di risoluzione temporale o spaziale. Ulteriori problemi possono derivare dagli elevati costi. Inoltre i pluviometri sono distribuiti in modo non uniforme e spesso posizionati piuttosto lontano dai centri urbani, comportando criticità e discontinuità nel monitoraggio. In questo contesto, un grande potenziale è rappresentato dall'utilizzo di tecniche innovative per sviluppare sistemi inediti di monitoraggio a basso costo. Nonostante la diversità di scopi, metodi e campi epistemologici, la letteratura sugli effetti visivi della pioggia supporta l'idea di sensori di pioggia basati su telecamera, ma tende ad essere specifica per dispositivo scelto. La presente tesi punta a indagare l'uso di dispositivi fotografici facilmente reperibili come rilevatori-misuratori di pioggia, per sviluppare una fitta rete di sensori a basso costo a supporto dei metodi tradizionali con una soluzione rapida incorporabile in dispositivi intelligenti. A differenza dei lavori esistenti, lo studio si concentra sulla massimizzazione del numero di fonti di immagini (smartphone, telecamere di sorveglianza generiche, telecamere da cruscotto, webcam, telecamere digitali, ecc.). Ciò comprende casi in cui non sia possibile regolare i parametri fotografici o ottenere scatti in timeline o video. Utilizzando un approccio di Deep Learning, la caratterizzazione delle precipitazioni può essere ottenuta attraverso l'analisi degli aspetti percettivi che determinano se e come una fotografia rappresenti una condizione di pioggia. Il primo scenario di interesse per l'apprendimento supervisionato è una classificazione binaria; l'output binario (presenza o assenza di pioggia) consente la rilevazione della presenza di precipitazione: gli apparecchi fotografici fungono da rivelatori di pioggia. Analogamente, il secondo scenario di interesse è una classificazione multi-classe; l'output multi-classe descrive un intervallo di intensità delle precipitazioni quasi istantanee: le fotocamere fungono da misuratori di pioggia. Utilizzando tecniche di Transfer Learning con reti neurali convoluzionali, i modelli sviluppati sono stati compilati, addestrati, convalidati e testati. La preparazione dei classificatori ha incluso la preparazione di un set di dati adeguato con impostazioni verosimili e non vincolate: dati aperti, diversi dati di proprietà del National Research Institute for Earth Science and Disaster Prevention - NIED (telecamere dashboard in Giappone accoppiate con dati radar multiparametrici ad alta precisione) e attività sperimentali condotte nel simulatore di pioggia su larga scala del NIED. I risultati sono stati applicati a uno scenario reale, con la sperimentazione attraverso una telecamera di sorveglianza preesistente che utilizza la connettività 5G fornita da Telecom Italia S.p.A. nella città di Matera (Italia). L'analisi si è svolta su più livelli, fornendo una panoramica sulle questioni relative al paradigma del rischio di alluvione in ambito urbano e questioni territoriali specifiche inerenti al caso di studio. Queste ultime includono diversi aspetti del contesto, l'importante ruolo delle piogge dal guidare l'evoluzione millenaria della morfologia urbana alla determinazione delle criticità attuali, oltre ad alcune componenti di un prototipo Web per la comunicazione del rischio alluvionale su scala locale. I risultati ottenuti e l'implementazione del modello corroborano la possibilità che le tecnologie a basso costo e le capacità locali possano aiutare a caratterizzare la forzante pluviometrica a supporto dei sistemi di allerta precoce basati sull'identificazione di uno stato meteorologico significativo. Il modello binario ha raggiunto un'accuratezza e un F1-score di 85,28% e 0,86 per il set di test e di 83,35% e 0,82 per l'implementazione nel caso di studio. Il modello multi-classe ha raggiunto un'accuratezza media e F1-score medio (macro-average) di 77,71% e 0,73 per il classificatore a 6 vie e 78,05% e 0,81 per quello a 5 classi. Le prestazioni migliori sono state ottenute nelle classi relative a forti precipitazioni e assenza di pioggia, mentre le previsioni errate sono legate a precipitazioni meno estreme. Il metodo proposto richiede requisiti operativi limitati, può essere implementato facilmente e rapidamente in casi d'uso reali, sfruttando dispositivi preesistenti con un uso parsimonioso di risorse economiche e computazionali. La classificazione può essere eseguita su singole fotografie scattate in condizioni disparate da dispositivi di acquisizione di uso comune, ovvero da telecamere statiche o in movimento senza regolazione dei parametri. Questo approccio potrebbe essere particolarmente utile nelle aree urbane in cui i metodi di misurazione come i pluviometri incontrano difficoltà di installazione o limitazioni operative o in contesti in cui non sono disponibili dati di telerilevamento o radar. Il sistema non si adatta a scene che sono fuorvianti anche per la percezione visiva umana. I limiti attuali risiedono nelle approssimazioni intrinseche negli output. Per colmare le lacune evidenti e migliorare l'accuratezza della previsione dell'intensità di precipitazione, sarebbe possibile un'ulteriore raccolta di dati. Sviluppi futuri potrebbero riguardare l'integrazione con ulteriori esperimenti in campo e dati da crowdsourcing, per promuovere comunicazione, partecipazione e dialogo aumentando la resilienza attraverso consapevolezza pubblica e impegno civico in una concezione di comunità smart.

Chaque année, le marché des transferts fait la une des journaux à cause des prix astronomiques payés par les grands clubs de football pour s'acheter des stars. Le montant payé par le club est supposé être une estimation de la valeur du joueur sur le marché. L'objectif de cette thèse est donc de déterminer les facteurs significatifs impactant le modèle de valorisation des joueurs. Pour ce faire, nous exploitons une base de données contenant 87 000 transferts et plus de 200 000 salaires avec deux types de variables ; une variable contenant des données statistiques sur chaque joueur pour les deux saisons précédentes, l'autre contenant une synthèse de notes données par les experts. Ce travail a été réalisé à partir d'un modèle de valorisation hédonique et de trois algorithmes de Machine Learning pour estimer les facteurs les plus importants dans la détermination de la valeur d'un joueur. Bien que perfectible, ces modèles sont capables de prédire les fonctions de prix des transferts et des salaires associés. Enfin, un modèle de marché a été implémenté pour déterminer l'effet des transferts, des résultats inattendus de matchs et de la Covid-19 dans la valeur d'un club de football. Ces recherches ont permis de fournir des explications prometteuses à propos des différentes segmentations sur le marché des transferts et l'impact de ces derniers sur la fluctuation de la valeur de certains clubs
Each year new transfer market news tops headlines due to the astronomical prices paid to recruit a superstar by top football clubs. The money paid by the buying club is assumed to be an estimate of the market value of the transferred player. Thus, the challenge is to determine the significant factors that affect the pricing function of a football player. In this research, a large data set has been extracted containing more than 87,000 transfers and more than 200,000 wage observation alongside two sets of variables; one contains real statistics of each player from the previous two seasons, while the other contains synthetic scores given by experts. This work has made use of one hedonic pricing function and three machine learning algorithms to estimate the most important factors affecting the financial value of the player. Albeit imperfect, but the models can predict the pricing functions of the transfer fees and wages with different promising precisions. Finally, a market model has been carried out to determine the effect of transfers, surprising match results, and COVID-19 on the market value of a football club. The overall findings were promising as they have provided interesting explanations about the different segmentations in the transfer market and the effectivity of transfers on the fluctuations of the share values of certain clubs

This thesis is a compilation of 3 articles, and the main focus of attention is the learning transfer mechanism. The first article explains the Rugby Attack Assessment instrument, a tool that assesses game performance of rugby union during a 5v5 small-sided game, taking into account simple actions and more complex tactical behaviours. The second paper explores the use of the Integrated Technique-Tactical Model used during the teaching and learning sequence, and reports the learning outcomes of this sequence. Results at a macro-level show no significant improvements. Results at a micro-level show an increase of some tactical behaviours frequencies. The third article analyses the interactivity among participants and the transfer of learning responsibility from the coach to the players. The units of analysis are the segments of interactivity. Results show that this process is linked to a slight decrease of segmentation, and mainly to the transfer of reflection from specific segments of discussion to reflections done during the guided practice
Aquesta tesi és una compilació de 3 articles, i l'objectiu principal és eñ mecanisme de traspàs de l'aprenentatge. El primer article explica el Rugby Attack Assessment Instrument, una eina que avalua el rendiment col·lectiuen el rugbi en una situació reduïda de 5x5, tenint en compte accions simples i comportaments tàctics més complexos. El segon article explica l'ús del Model Integrat Tècnic-Tàctic utilitzat durant la seqüència d'ensenyament i aprenentatge, i explica els resultats de l'aprenentatge d'aquesta seqüència. Els resultats en un nivell macro revelen que no hi ha millores significatives. Els resultats a nivell micro mostren un increment de la freqüència de determinats comportaments tàctics. El tercer article analitza la interactivitat entre els participants i el traspàs de la responsabilitat de l'aprenentatge de l'entrenador als jugadors. Les unitats d'anàlisi són els segments d'interactivitat. Els resultats mostren que aquest procés està lligat a un lleuger descens de la segmentació i principalment a un traspàs dels moments de reflexió des de segments específics de discussió cap a reflexions dutes a terme durant la pràctica guiada

Over the past decade, Australian universities have experienced an exponential increase in the enrolment of fee-paying overseas students whose preparation for tertiary studies may differ significantly from that of local students. Despite English language proficiency requirements, there is some concern that international entry tests do not adequately measure the complex features of university writing; an important concern given that student success is heavily dependent on their mastery of academic writing. As a result, many international students require additional support structures. Until the present, debate about the most effective way to meet the diverse needs of English as an Additional Language (EAL) writers entering universities has concerned a choice between two alternatives: on one hand a separate, short-term English for Academic Purposes (EAP) language program and on the other, direct entry into disciplines with lecturers taking responsibility for assisting students to learn the discipline-specific language skills required. While the Australian Universities Quality Agency (AUQA, 2009, 2013) supports the latter view, this research investigates a third alternative; that is, an English for Academic Purposes Pathway program (EAPP) that not only teaches general academic English skills, but also English required in discipline specific contexts, as well as important and necessary adjunct skills that support writing. This three-phase, mixed-methods study used both qualitative and quantitative data to investigate the efficacy of such a program. The study, which was analytic, descriptive and comparative in approach, was conducted in a naturalistic setting and, where possible, qualitative data were used to support the findings from quantitative data. Theoretical propositions guided the data collection and provided important links to connect primary and secondary research. Phase 1 investigated the academic writing needs perceived by 60 students who were either studying in the 20-week or 10-week EAPP program at Swan University (a pseudonym). Perceptions of student needs by 13 EAPP teachers were also analysed and writing samples collected. In Phase 2, the cohort decreased to 31 students representing seven faculties. Perceptions of 17 faculty staff from across and within these seven faculties were sought regarding the tasks and genres required for EAL students to meet the writing expectations within these disciplines. The marked ex-EAPP student’s faculty writing assignments were collected and analysed at the end of first semester. At this stage, because the volume of student writing produced over the course of the study was so large, disproportional stratified random sampling was used to select and analyse the EAPP and faculty writing of a sample of seven students. Research by Kaldor, Herriman and Rochecouste (1998) provided direction for frame analysis which was used to analyse the student writing. In Phase 3, which was conducted one year after entering their chosen faculties, 22 students replied to a request to judge which, if any, writing skills from their EAPP program had transferred to assist them with their faculty writing. Findings are discussed in relation to four major issues. Firstly, reflections provided by ex-EAPP students ascertained that, on entering the EAPP program, the majority of them had been academically, linguistically, culturally and socially unprepared for study at master’s degree level in an Australian university. Secondly, analysis determined that in the students’ first year of faculty study, writing tasks and genres were almost identical in type, complexity and word-count restrictions to those taught in the EAPP program and that students readily adapted to the highly specified frameworks of any tasks that were unfamiliar. A third major finding was the significance that students placed on the type of feedback necessary to support their writing. Finally, students identified major areas of improvement in their academic writing at the end of the program, but provided suggestions in key pedagogical areas about how the EAPP program could be improved to better address their needs. This study found that EAL writing development involves much more than content knowledge, mastery over discipline-specific genre requirements and a wide vocabulary. Academic writing comprises a complex combination of extratextual, circumtextual, intratextual and intertextual features and skills, some of which are completely new to international students. A model was proposed to illustrate elements that provide: circumtextual assistance for prewriting support; intertextual assistance through reading and writing support; extratextual assistance through sociocultural support, and intratextual assistance through the scaffolding of academic writing skills. To conclude, recommended modifications to the program are presented.

Applications of machine learning demand exploration of feature extraction methods and classifier types in order to obtain systems with reliable highest accuracy. The industrial application discussed in this thesis is the classification of ultrasonic echoes in an A-scan. The application is particularly challenging as A-scans are taken from the end of long large complex metal shafts. Although several pattern analysis and machine learning techniques have been used with success in analyzing A-scan data [43, 89], they are typically in the context of very short signals produced from machine parts like plate surfaces or pipe surfaces. Those cases are usually much simpler; in particular, the task reduces to detecting the existence of an echo (indicating a fault in the material). When signals for testing come from long shafts, however, a major problem of mode-converted reflection emerges. These reflections are echoes that do not correspond to real faults (cracks), nor to characteristics in the shaft. These mode-converted echoes may cause misjudgement of the position of cracks on shafts as some critical faint echoes from a cracked surface may lie somewhere among the multiple secondary echoes. Consequences of misclassification are catastrophic with enormous cost in downtime, consequential damage to associate equipment and potential injury to personnel [23]. The problem is then, to discriminate efficiently the different types of reflectors amongst the large volumes of digitalized ultrasonic shaft defect information. As the relationship between ultrasonic signal characteristics and flaw classes is not straightforward, we need to identify and extract informative sets of signal features from which classification might be performed more efficiently and accurately. Among various methods for extracting signal features, the Fast Fourier Transform (FFT) is a useful scheme for extracting frequency-domain signal features [23, 62]. This seems natural when dealing with ultrasound since the traditional representation of these types of signals is by mathematical Fourier series that identify physically meaningful features, like frequency and phase. But recent studies on the ultrasonic flaw classification employ the Discrete Wavelet Transform (DWT) as part of their feature extraction scheme. DWT provides effective signal compression and time-frequency presentation [69, 86]. Many researchers have compared these two feature extraction schemes (FFT and DWT), and most comparisons showed a superiority of DWT to FFT in discriminating the type of flaw (or its non-existence) [74, 78, 90]. However, these previous reported studies have compared the DWT based features with the FFT with limited feature components. Typically, short signals have been reported, with little attention to phase components of FFT sequences. This thesis is the first study analyzing feature extraction in more complex ultrasonic signals from shafts. In particular, we introduce a new FFT-based feature extraction scheme FFT_Magpha which effectively represents both magnitude and phase components of FFT sequences. By employing this state-of-the-art FFT feature extraction scheme, we have more extension and reliability in the investigation about the feasibility of FFT as a better feature extraction scheme than other types of feature extraction schemes such as DWT. On the other hand, the time-variance problem exhibited in DWT has resulted in reservations about its wide acceptance even though DWT coefficients provide effective time-frequency representation of non-stationary signals, and thus are considered useful features for input into classifiers. To solve this, we study a new preprocessing technique for time-domain A-scans, which offer consistent extraction of a segment of the signal from long signals that occur in the NDT of shafts. We compare the performance of this new echo-gating technique with other previously developed methods and investigate that we can use DWT more efficiently as a feature extraction scheme for ultrasonic signal classification by employing this new method in the preprocessing stage. In addition, our investigation in this thesis finds the potential of DWT to be a more reliable feature extraction scheme, through the more stable classification results in different runs of cross validation tests than the results produced in the tests using FFT-based feature extraction scheme. This potential is especially beneficial for the practical NDT for shafts, where we can train a classifier with arbitrary training data and then use the classifier for in-field ultrasonic shaft signal test. We also demonstrate the superiority of using DWT as the feature extraction scheme in the ultrasonic shaft signal classification involving not only ANN hut also SVM. These results dissipate any doubt that the DWT feature extraction methodology is too far suited for ANN which has been popularly employed previously in many similar experimental scenarios. Through these experimental comparisons employing various learning algorithms, we find a certain facility when specific classes of echoes are concerned with different combinations of feature extraction (FFT or DWT) and classifier (ANN or SVM), though DWT is superior to FFT and SVM is superior to ANN in terms of the overall classification accuracy. This finding leads into a hybrid classifier that will improve overall performance by giving more weight to the more trustworthy sub-classifier. Based on those experimental analysis, we design an Integrated SVM classifier (ISVM) which is a combined classification system efficiently employing benefits from each of two SVM classifiers using two feature extraction schemes, FFT aud DWT. The outcomes of a classifier based on FFT is not totally dismissed in this system although the DWT-based classifier has been shown to be superior. This property of ISVM enables us to combine classifiers considering the misclassification cost, to obtain a more informative classification for its application in the field. We also explore the diverse possibilities of heterogeneous and homogeneous ensembles by combining the classifiers along the dimension of feature extraction mechanism, along the dimension of combination method and along the dimension of type of classifier. The experimental result suggests guidelines for designing an integrated multi-classifier system for shaft test data by way of selectively employing the combining structure.
Thesis (PhD Doctorate)
Doctor of Philosophy (PhD)
School of Information and Communication Technology
Full Text

Iskustvo u radu sa studentima pokazalo je da studenti nedovoljno ĉesto ostvaruju zadovoljavajuće ishode uĉenja na lingvistiĉkim predmetima na studijama engleskog jezika, što je takoĊe opisano i u stranoj literaturi. Kao jedan od razloga navodi se nastava koja upućuje studente na pasivnost, te oni pribegavaju memorisanju i reprodukciji materijala kao najĉešćim ishodima uĉenja. Biggs (1999) razraĊuje koncepte dubinskog i površinskog pristupa uĉenju, koje relevantna literatura smatra kljuĉnim faktorima za kvalitet ishoda uĉenja. Dubinski pristup uĉenju korelira sa kvalitetnim ishodima uĉenja i funkcionalnim znanjem, a karakterišu ga motivacija, zadovoljstvo usled uĉenja, studentska aktivnost i to aktivnost visokih kognitivnih funkcija. Visoke kognitivne funkcije (Bloom i dr. 1956, Anderson i dr. 2001) i povezane kognitivne radnje (rešavanje problema, analitiĉko, kritiĉko i kreativno razmišljanje) jesu najvaţniji ciljevi visokog obrazovanja jer samo one, usled primene transfera uĉenja, govore o steĉenom i primenljivom, tj. funkcionalnom znanju. Obuka i razvoj visokih kognitivnih funkcija omogući će studentima da uĉenju pristupe dubinski što je još jedan razlog da budu osnovni nastavni cilj svih predmeta na ustanovama tercijarnog obrazovanja. U skladu sa ovim teorijskim postavkama postavljene su osnovna i pomoćna hipoteza: upotreba posebno konstruisanih veţbi za aktivaciju viših kognitivnih funkcija u nastavi lingvistiĉkih predmeta dovešće do sticanja funkcionalnog znanja na teorijskom i praktiĉnom nivou; steĉeno znanje kroz ovakvu eksperimentalnu nastavu i upotreba tog znanja odraţavaće kognitivne funkcije ne samo niţeg nego i višeg reda: primeniti, analizirati, proceniti, stvoriti, kao i kritiĉko i kreativno razmišljanje i rešavanje problema. Kako bi se proverile hipoteze, sproveden je eksperiment sa studentima prve godine engleskog jezika (N=34) na Fakultetu za pravne i poslovne studije dr Lazar Vrkatić u Novom Sadu. U istraţivanju sa paralelnim grupama, eksperimentalna grupa je imala veţbe sa aktivnostima koje razvijaju više kognitivne funkcije na predmetu uvod u opštu lingvistiku tokom zimskog semestra školske 2012/2013. godine. UporeĊeni su kvantitativni rezultati kolokvijuma eksperimentalne i kontrolne grupe na kraju semestra, a potom je sproveden intervju sa po pet studenata iz svake grupe radi utvrĊivanja kvalitativnih razlika u kognitivnim procesima kod ove dve grupe. Obe grupe su ostvarile podjednak uspeh na kolokvijumu, te je osnovna hipoteza odbaĉena. Kodirani podaci iz intervjua pokazali su da obe grupe podjednako koriste kognitivne funkcije po broju i distribuciji, te je i pomoćna hipoteza odbaĉena. Kao objašnjenje za odsustvo većeg uspeha EG navedena su metodološka ograniĉenja istraţivanja: duţina eksperimentalne nastave, problem dokazivosti transfera i problem kodiranja intervjua. Drugi faktori koji mogu objasniti neuspeh su: prethodno steĉene navike u uĉenju, neshvatanje svrhe izuĉavanja predmeta i dr. UporeĊeni su rezultati boljih i slabijih studenata, te je utvrĊeno da bolji studenti pokazuju veći stepen samostalnosti, da upotrebljavaju više kognitivne funkcije kao i duţe nizove kognitivnih radnji. Posmatrajući kvalitativne podatke, bolji studenti eksperimentalne grupe pokazali su promenu gledanja na svet usled izuĉavanja lingvistike i izrazili su zadovoljstvo zbog uĉenja ovog predmeta. Oni pokazuju i upotrebu najduţih nizova vezanih kognitivnih radnji. Iz ovoga se moţe zakljuĉiti da su oni pristupili uĉenju dubinski i zbog toga ostvarili kvalitetnije ishode uĉenja. U cilju donošenja ĉvrstih zakljuĉaka neophodno je sprovesti dugotrajniji i obuhvatniji multidiciplinarni istraţivaĉki projekat, s obzirom da bi pozitivni rezultati bili od velikog znaĉaja za poboljšanje ishoda uĉenja na tercijarnom nivou obrazovanja. Ključne reči: uĉenje i nastava na tercijarnom nivou, taksonomija obrazovnih ciljeva, pristupi uĉenju, više kognitivne funkcije, transfer uĉenja, funckionalno znanje.
The experience of working with university students has shown that the learning outcomes of linguistic courses are infrequently satisfactory, which is also described in literature worldwide. Teaching philosophy in which students are forced into passives roles is one of the causes since such teaching results in low motivation with memorizing and reproduction of learned materials as the most frequent outcomes of learning. Biggs (1999) develops the concepts of deep and superficial learning approaches which have been declared in the relevant literature as key factors for the quality of learning outcomes. Deep approach to learning correlates with high quality learning outcomes, and is characterized by high motivation, satisfaction with learning and student activity of appropriately high cognitive levels. Higher cognitive functions (Bloom et. al. 1956, Anderson at al. 2001) and related cognitive activities (problem solving, analytical, critical and creative thinking) are the most important goals of higher education since these thinking skills are transferable and therefore represent applicable and functional knowledge. The training and development of the higher cognitive skills enables students to use deep approaches to learning, which is an additional reason to consider them as fundamental teaching goals in all courses in tertiary education. Based on this theoretical framework the main hypothesis and sub-hypothesis were formulated as follows: the use of specially designed practices which activate higher cognitive functions (HCF) will result in acquiring functional knowledge at both theoretical and practical levels; the knowledge gained through such teaching will reflect the use of higher cognitive functions: apply, analyze, evaluate, create, as well as show problem solving skills and critical and creative thinking. To test the hypotheses an experiment was conducted with the first year English language students (N=34) at the Faculty of Legal and Business Studies dr Lazar Vrkatić in Novi Sad. In the parallel groups design, the experimental group (EG) was involved with activities which develop HCFs in the course of Introduction to General Linguistics during the winter semester of the 2012/2013. Quantitative data were collected at the end of the semester (the final test) and compared between the two groups to determine whether the EG scored better results than the control group (CG). This was followed by interviews with five respondents from each group to qualitatively compare the cognitive processes. No statistically significant difference between test results in the two groups was found and so the main hypothesis was rejected. The coded data from the interviews showed an equal number of identified CFs with both groups with similar distribution patterns, thus the sub-hypothesis was also rejected. The absence of better scores of the EG can be explained by some methodological limitations of the experiment, such as the length of the experimental activities, the problem of proof of transfer and the coding of the interview data. Other factors include the existing learning habits of students, the inability to grasp the purpose of studying linguistics, etc. The results of better students were compared to those of the weaker ones, which showed that better students are more autonomous, use a greater number of HCFs and string more CFs into a complex response. Qualitative data also showed that better students of the experimental group expressed a change in how they see the world around them and express satisfaction because of studying linguistics. They also string the longest chains of cognitive activities. These findings lead to a conclusion that better students of the EG used deep approaches to learning which resulted in higher quality learning outcomes. In order to achieve conclusive results, a comprehensive long-term multidisciplinary research project should be carried out, since its results would have a significant impact on the quality of learning outcomes in tertiary education.

Cette recherche doctorale s’inscrit dans le cadre de la didactique de l’anglais et prend appui sur les champs théoriques propres à la didactique des langues et aux sciences du langage. Elle a pour objet l’étude de l’introduction d’approches plurielles fondées sur les langues inconnues (APLI) au sein du cours d’anglais et cherche à comprendre les effets d’une telle modification du savoir sur les acteurs de la relation pédagogique (l’élève et l’enseignant). Une quasi-expérimentation fut menée sur l’année scolaire 2011-2012 dans cinq classes de collège, de niveau cinquième. Les élèves, regroupés par tétrades, ont successivement été confrontés à trois langues inconnues (néerlandais, italien et finnois). Pour chacune de ces langues, ils ont été amenés à résoudre des problèmes d’ordre métasémantique, métasyntaxique, puis métaphonologique. Après avoir circonscrit le cadre institutionnel et théorique nécessaire à l’étude, le cadre méthodologique est précisé. Puis est abordée l’analyse des effets de la modification du savoir devenu plurilingue, tant sur les élèves que sur les enseignants. Lorsqu’on observe les effets des APLI sur la relation Savoir-Professeur, on constate que cette modification didactique a permis aux enseignants de mieux comprendre les concepts de situation-problème, de conceptualisation, de stratégies d’apprentissage et de compétence. La transposition didactique s’en trouve modifiée : les enseignants ont progressivement été amenés à concevoir des séquences didactiques donnant plus de place à l’élève, avec des exigences plus élevées. L’étude de la relation Enseignant-Élève met en exergue une modification de la pratique, essentiellement liée à la mise en place du travail de groupe. Le rôle de l’enseignant est alors révisé : il devient facilitateur du travail qui s’effectue en collaboration au sein du groupe. Finalement, l’analyse de la relation Savoir-Élève souligne la nécessaire conscientisation qui mène vers le développement de compétences plurilingues et la mise en œuvre de stratégies d’apprentissage transférables à l’étude de la L2
This doctoral research work is embedded in the field of language didactics and is equally based on the linguistics and cognitive theoretical fields. Its main goal is to study the introduction of pluralistic approaches based on unknown languages (PAUL) within the English class, at lower secondary school. It seeks to understand the effects of such a change of knowledge on the actors of the pedagogical relationship (student and teacher). A quasi-experiment was conducted in 2011-2012 in five year 7 and four year 9 forms. Students, in groups of four, were successively confronted to three unknown languages (Dutch, Italian and Finnish). They were asked to solve metasemantic, metasyntactic or metaphonological problems in turn, for each of these languages. This doctoral work first explores the institutional and theoretical framework. Then, it presents the methodological framework so as to be able to analyze the effects of the change of the knowledge parameter which has become multilingual, both on the students and the teachers. When looking into the effects of PAUL on the Knowledge-Teacher relationship, we observe that it enables teachers to better apprehend concepts such as problem-solving, conceptualisation, learning strategies and competence. The didactic transposition is hence modified: teachers have gradually been led to develop teaching sequences with higher standards giving more space to the student. The study of the Teacher-Student relationship highlights a change in practice, mainly due to the implementation of group work. The role of the teacher is then revised: he becomes a facilitator of the collaborative learning. Finally, the analysis of the Knowledge-Student relationship underlines the necessary awareness that leads to the development of multilingual competences through the implementation of learning strategies which appear to be transferable to the study of L2

The poor state and failure of the basic education system in South Africa gave rise to this research. The wave of knowledge loss experienced in the last two decades is expected to carry on and will continue to deplete the basic education system’s knowledge base, severely affecting the already poor quality of education as well as the future economic growth and sustainability in South Africa. The main research objective was to establish whether future growth and sustainability in the basic education system in South Africa is achievable; which factors it is influenced by; and how knowledge continuity could impact on future growth and sustainability. A multidisciplinary approach focusing on organisational performance, knowledge management, individual and organisational behaviour and organisational development was followed. The nature of growth and sustainability and knowledge continuity in organisations was explored by following a contextualisation theory-building process. The main objective of the empirical research study was to determine by means of quantitative research the degree to which the influencing factors would enhance or impede growth and sustainability in an organisation. A quantitative survey method was followed. A questionnaire was developed and the survey was performed in 6 primary and secondary schools of the basic education system in South Africa. The questionnaire was found to be reliable with a Cronbach’s alpha of .8060. In the descriptive factor analysis process, principal component factor analysis was conducted, which described the five constructs that would influence growth and sustainability. These constructs’ dimensions produced significant intercorrelations which indicate that the dimensions are for the most part intercorrelated with each other in contributing to growth and sustainability. The multiple regression analysis indicated that knowledge loss would have an exceptionally strong impact on knowledge; and that knowledge, information and performance would significantly predict growth and sustainability. Organisations should change the focus for growth from physical assets to the development of intellectual capital, and knowledge continuity should form part of an organisations’ business strategy and mission. Knowledge continuity will only be successful if a culture conducive of trust and knowledge sharing and transfer exist, and are supported by effective and appropriate human resource practices and incentives. A structural equation model development strategy produced a knowledge continuity model aimed at enabling future growth and sustainability, based on the constructs confirmed in the factor analysis. The model indicated that there is a direct causal relationship between knowledge, information and performance with growth and sustainability. The regression analysis showed that most of the intercorrelations are significant, thus confirming the theory. The newly developed questionnaire and structural equation model should enable organisations to measure the degree to which the enhancing individual and organisational behavioural factors of growth and sustainability are in place and provide the measurement outcomes that would identify the factors that need to be focused on to improve and enable future growth and sustainability in an organisation.
Thesis (MBA)--North-West University, Potchefstroom Campus, 2013.

碩士
淡江大學
電機工程學系碩士班
106
Machine learning and deep learning particularly have gained a lot of attention in recent years, especially for classification related tasks, such as text mining, face and speech, etc. The performance increase is mostly due to complex algorithm and architecture, and partly due to the use of good data sets. The main motivation of this thesis is to train a Convolutional Neural Network (CNN) based system for face recognition aiming at positive prediction and appreciative accuracy result. By way of transfer learning, a pre-trained model can be tailored for different applications with new data. The resulting output attains good accuracy and result in different cases. The objective is to differentiate 3 labeled categories, each with 200 images in the training dataset. The training data is provided to modify the pre-trained model, which is further classified with the test images in different scenarios, where the prediction results achieve high accuracy for each individual case.

Esta dissertação tem como objectivo analisar a promoção da aprendizagem activa nas secções de formação que utilizem digital game-based learning e em que medida essa promoção se traduz num aumento da adaptabilidade dos formandos. O estudo consistiu num questionário que mediu as percepções que os jogadores de jogos de vídeo tinham acerca da presença de componentes de treino (aprendizagem exploratório e gestão de erros), processos auto-regulatórios (actividade metacognitiva, motivação intrínseca e auto-eficácia) e de resultados de aprendizagem (transferência analógica e transferência adaptável) em jogos de vídeo. O questionário foi preenchido por uma amostra de 220 indivíduos de diversas idades. Os resultados do estudo mostraram que quando jogos de vídeo são percepcionados como indutores de aprendizagem exploratória ou gestão de erros, têm uma relação positiva com os diversos processos auto-regulatórios. Para além disso, os processos auto-regulatórios actuam como mediadores da relação entre os componentes de treino e os resultados da aprendizagem testados. O estudo também descobriu que reacções emocionais negativas a erros têm um impacto positivo fraco nas actividades metacognitivas e uma relação positiva com a transferência adaptável, mediada pela actividade metacognitiva. As conclusões apresentadas permitem considerar os jogos de vídeo como adições relevantes a formação organizacional profissional, uma vez que estes têm o potencial de promoverem os mesmos resultados de aprendizagem que outras intervenções que usaram uma abordagem de aprendizagem activa, quando incluem os componentes de treino requeridos. O presente trabalho é o primeiro que tenta analisar digital game-based learning como uma abordagem capaz de promover componentes de aprendizagem activa e, igualmente importante, o primeiro que analisa a capacidade que os jogos de vídeo têm de promover a transferência adaptável dos conceitos apreendidos enquanto os jogam.
This dissertation aims to analyse the promotion of active learning components in training interventions that use digital game-based learning and, moreover, if said promotion translates in an increase of the adaptability of the trainees. The study consisted on a survey that measured the perceptions of video game players about the training components (exploratory learning and error framing), self-regulatory processes(metacognitive activity, intrinsic motivation and self-efficacy) and the learning outcomes (analogical transfer and adaptive transfer) promoted by playing the games. The survey was answered by 220 persons of several ages. The results of the study showed that when video games were perceived as inducing exploratory learning or error framing they had positive relationships with several self-regulatory processes. Also, these self-regulatory processes also acted as mediators in the relation between the training components and the learning outcomes tested. The study also discovered that a negative emotional reaction to errors had a weak positive impact on the metacognitive activity, and had a positive relation with adaptive transfer, mediated by metacognitive activity. The presented conclusions lead to the consideration that video games are relevant additions to professional training in organizations, as they have the potential of promoting the same learning outcomes present in training interventions that used an active learning approach, if embedded with the required training outcomes. The present work it is the first one, to the knowledge of the researchers, to see analyse digital game-based learning as an approach capable of promoting an active learning intervention and equally important the first to analyse the capacity that video games have of promoting an adaptive transfer of the learned concepts while playing.

碩士
國立臺灣科技大學
資訊工程系
100
We proposed a method to deal with the classification of high intra- class variation data based on a transfer learning approach. The high intra-class variation is difficult to model especially when we only have a limited dataset. In this case, a single concept may consist of several diverse sub-concepts and each concept has only very few samples. The boosting or Adaboost, for instance, can not help much in this case be- cause we may easily produce a weak classifier that gives error rate higher than one half and as a result, the boosting procedure will halt. We pro- pose a transfer learning approach to effectively integrate the information from high-variation samples for a successful modeling. In our approach, we put samples of high variation into the source and target domains, as in the design of TrAdaboost; then gradually, we select some useful data from the source domain and combine them with the data in the target domain to form a rich set for training. What is different from the TrAdaboost is that in our approach, the weight of data in the source domain is not necessarily decreased as always; therefore, we can collect more useful data from the source domain based on the proposed method than based on the typical TrAdaboost. Our contribution is twofold: on one hand, we can successfully deal with high intra-class variation data; on the other hand, we can also improve the performance of TrAdaboost, when the data in the source and target domains are with high variation. The experiment result shows that the proposed method can achieve higher accuracy than that of other classification method such as Adaboost for the classification ofhigh intra-class variation data; moreover, the proposed method performs better than that of TrAdaboost for the same types of data.

碩士
國立臺灣大學
資訊工程學研究所
101
Generally qualitative condition (the accuracy of the data) and quantitative condition (the amount of data) of the data can significantly affect the quality of a supervised learning model. However, in real-world applications it might not be feasible to always assume one can obtain large amount of high-quality datasets. This research assumes the situation that there is a only small amount of accurate training data available for learning, aiming at designing a transfer-learning based approach to utilize larger amount of noisy (in terms of labels and features) training data to improve the learning quality. This problem is non-trivial because the distribution in noisy training dataset is different from that of the testing data. In this thesis, we proposed a novel transfer learning algorithm, Noise-Label Transfer Learning (NLTL), to solve the problem. We exploit the information of labels and features from accurate and noise data, transferring the features into same domain and adjusting the weights of instances for learning. The experiment result shows NLTL could outperform the existing approaches.

碩士
元智大學
資訊管理學系
107
In recent years, deep learning technology has been highly developed in image recognition, and is also widely used in natural language recognition and word exploration. This study uses migration learning techniques to analyze online reviews and then extract and amplify keywords through feature engineering. This study uses deep learning techniques to load the migration learning mechanism and text classification study. This study will add an Attention Layer into general deep neural network, then through combining multiple deep neural networks and Stacking technology, a final model is developed as for comments detection. The experimental results of detect the extreme comments show that using the deep neural network with Attention Layer, the detection results can be 66.19% in F1 measure and Auc: 96.05%. The combined deep neural network with Stacking technology approach can obtain F1 measure 69.96% and Auc: 96.17%. This study involved Kaggle nature language competition of extreme contents detection on Quora. The results of this study ranked within the top 16% of the global competition, F1 measure 70.13%, and the best winning result of 71.32%, with only 1.2% difference.

The prevalence of deepfake technology has led to serious worries about the veracity and dependability of visual media. To reduce any harm brought on by the malicious use of this technology, it is essential to identify deepfakes. By using the Vision Transformer (ViT) model for classification and the InceptionResNetV2 architecture for feature extraction, we offer a novel approach to deepfake detection in this thesis. The highly discriminative features are extracted from the input photos using the InceptionResNetV2 network, which has been pre-trained on a substantial dataset. The Vision Transformer model then receives these characteristics and uses the self attention method to identify long-range relationships and categorize the pictures as deepfakes or real. We use transfer learning techniques to improve the performance of the deepfake detection system. The InceptionResNetV2 model is fine-tuned using a deep fake specific dataset, which allows the pre-trained weights to adapt to whatever task is at hand, allowing the extraction of meaningful and discriminative deepfake features. Following that, the refined features are put into the ViT model for categorization. Extensive experiments are conducted to evaluate the performance of our proposed approach using various deepfake datasets. The results demonstrate the effectiveness of the InceptionResNetV2 and ViT combination, achieving high accuracy and robustness in deepfake detection across different types of manipulations, including face swapping and facial re-enactment. Additionally, the utilization of transfer learning significantly reduces the training time and computational resources required to train the deepfake detection system. This research's outcomes contribute to advancing deepfake detection techniques by leveraging state-of-the-art architectures for feature extraction and classification. The fusion of InceptionResNetV2 and ViT, along with the implementation of transfer learning, offers a powerful and efficient solution for accurate deepfake detection, thereby safeguarding the integrity and trustworthiness of visual media in an era of increasing digital manipulation.

Indiana University-Purdue University Indianapolis (IUPUI)
Deep learning solutions to computer vision tasks have revolutionized many industries in recent years, but embedded systems have too many restrictions to take advantage of current state-of-the-art configurations. Typical embedded processor hardware configurations must meet very low power and memory constraints to maintain small and lightweight packaging, and the architectures of the current best deep learning models are too computationally-intensive for these hardware configurations. Current research shows that convolutional neural networks (CNNs) can be deployed with a few architectural modifications on Field-Programmable Gate Arrays (FPGAs) resulting in minimal loss of accuracy, similar or decreased processing speeds, and lower power consumption when compared to general-purpose Central Processing Units (CPUs) and Graphics Processing Units (GPUs). This research contributes further to these findings with the FPGA implementation of a YOLOv4 object detection model that was developed with the use of transfer learning. The transfer-learned model uses the weights of a model pre-trained on the MS-COCO dataset as a starting point then fine-tunes only the output layers for detection on more specific objects of five classes. The model architecture was then modified slightly for compatibility with the FPGA hardware using techniques such as weight quantization and replacing unsupported activation layer types. The model was deployed on three different hardware setups (CPU, GPU, FPGA) for inference on a test set of 100 images. It was found that the FPGA was able to achieve real-time inference speeds of 33.77 frames-per-second, a speedup of 7.74 frames-per-second when compared to GPU deployment. The model also consumed 96% less power than a GPU configuration with only approximately 4% average loss in accuracy across all 5 classes. The results are even more striking when compared to CPU deployment, with 131.7-times speedup in inference throughput. CPUs have long since been outperformed by GPUs for deep learning applications but are used in most embedded systems. These results further illustrate the advantages of FPGAs for deep learning inference on embedded systems even when transfer learning is used for an efficient end-to-end deployment process. This work advances current state-of-the-art with the implementation of a YOLOv4 object detection model developed with transfer learning for FPGA deployment.

Deep learning solutions to computer vision tasks have revolutionized many industries in recent years, but embedded systems have too many restrictions to take advantage of current state-of-the-art configurations. Typical embedded processor hardware configurations must meet very low power and memory constraints to maintain small and lightweight packaging, and the architectures of the current best deep learning models are too computationally intensive for these hardware configurations. Current research shows that convolutional neural networks (CNNs) can be deployed with a few architectural modifications on Field-Programmable Gate Arrays (FPGAs) resulting in minimal loss of accuracy, similar or decreased processing speeds, and lower power consumption when compared to general-purpose Central Processing Units (CPUs) and Graphics Processing Units (GPUs). This research contributes further to these findings with the FPGA implementation of a YOLOv4 object detection model that was developed with the use of transfer learning. The transfer-learned model uses the weights of a model pre-trained on the MS-COCO dataset as a starting point then fine-tunes only the output layers for detection on more specific objects of five classes. The model architecture was then modified slightly for compatibility with the FPGA hardware using techniques such as weight quantization and replacing unsupported activation layer types. The model was deployed on three different hardware setups (CPU, GPU, FPGA) for inference on a test set of images. It was found that the FPGA was able to achieve real-time inference speeds of 33.77 frames-per-second, a speedup of 7.74 frames-per-second when compared to GPU deployment. The model also consumed 96% less power than a GPU configuration with only approximately 4% average loss in accuracy across all 5 classes. The results are even more striking when compared to CPU deployment, with 131.7-times speedup in inference throughput. CPUs have long since been outperformed by GPUs for deep learning applications but are used in most embedded systems. These results further illustrate the advantages of FPGAs for deep learning inference on embedded systems even when transfer learning is used for an efficient end-to-end deployment process. This work advances current state-of-the-art with the implementation of a YOLOv4 object detection model developed with transfer learning for FPGA deployment.

A new, novel numerical optimization algorithm is developed, tested, and used to solve difficult numerical problems from the field of astrodynamics. First, a brief review of optimization theory is presented and common numerical optimization techniques are discussed. Then, the new method, called the Learning Approach to Sampling Optimization (LA) is presented. Simple, illustrative examples are given to further emphasize the simplicity and accuracy of the LA method. Benchmark functions in lower dimensions are studied and the LA is compared, in terms of performance, to widely used methods. Three classes of problems from astrodynamics are then solved. First, the N - impulse orbit transfer and rendezvous problems are solved by using the LA optimization technique along with derived bounds that make the problem computationally feasible. This marriage between analytical and numerical methods allows an answer to be found for an order of magnitude greater number of impulses than are currently published. Next, the N -impulse work is applied to design periodic close encounters (PCE) in space. The encounters are defined as an open rendezvous, meaning that two spacecraft must be at the same position at the same time, but their velocities are not necessarily equal. The PCE work is extended to include N -impulses and other constraints, and new examples are given. Finally, a trajectory optimization problem is solved using the LA algorithm and comparing performance with other methods based on two models-with varying complexity-of the Cassini-Huygens mission to Saturn. The results show that the LA consistently outperforms commonly used numerical optimization algorithms.

碩士
國立中央大學
資訊工程學系
107
Machine translation has been research for a long time. Although most of the sentences can be translated correctly, when it comes to named entity like a personal name or a location in a sentence, there's still room for improvement especially between non-English languages. Named Entity Transliteration is a way to solve the condition mentioned above. Transliteration is a key part of machine translation. However when we actually do research, we often have limited parallel data between source language and target language. If we take a wildly used language as a pivot langage, in contract, it would be more easily to extract language pairs of source language to pivot language and pivot language to target language. It's intuitive to extract the common pivot language entities from these corpora to generate a three-language parallel data include source language, pivot language, target language. We can achieve the bilingual transliteration task using the parallel data; nevertheless, large amount of data is wasted in this method. We propose a modified attention-based sequence-to-sequence model which also applies transfer learning techniques. Our model effectively utilize the remaining data besides the parallel data to promote the performance of named entity transliteration.

碩士
國立臺北大學
資訊工程學系
107
The frequency of pest occurrence has always been a task of agricultural time and labor. This paper attempts to solve the above problems through the combination of deep learning and agriculture. We propose an entire-and-partial feature transfer learning scheme to perform pest detection, classification and counting, to offer the result of pest occurrence frequency. In the partial-feature transfer learning, the fine-grained feature map of the partial-feature transfer learning is used to strengthened the entire-feature transfer learning. Finally, different fine-grained feature map are strengthened to the entire-feature transfer learning use weight scheme and the cross-layer of the entire-feature network is combined with multi-scale feature map. The entire-feature transfer learning approach enhances the feature by creating a shortcut topology using cross layer mechanism to reduce the gradient disappearance problem. The experimental results shows that the detection and classification of the entire-and partial feature transfer learning mechanism can be significantly improved, and the method can reach 90.2%.

Sleep occupies significant part of human life. The diagnoses of sleep related disorders are of great importance. To record specific physical and electrical activities of the brain and body, a multi-parameter test, called polysomnography (PSG), is normally used. The visual process of sleep stage classification is time consuming, subjective and costly. To improve the accuracy and efficiency of the sleep stage classification, automatic classification algorithms were developed. In this research work, we focused on pre-processing (filtering boundaries and de-noising algorithms) and classification steps of automatic sleep stage classification. The main motivation for this work was to develop a pre-processing and classification framework to clean the input EEG signal without manipulating the original data thus enhancing the learning stage of deep learning classifiers. For pre-processing EEG signals, a lossless adaptive artefact removal method was proposed. Rather than other works that used artificial noise, we used real EEG data contaminated with EOG and EMG for evaluating the proposed method. The proposed adaptive algorithm led to a significant enhancement in the overall classification accuracy. In the classification area, we evaluated the performance of the most common sleep stage classifiers using a comprehensive set of features extracted from PSG signals. Considering the challenges and limitations of conventional methods, we proposed two deep learning-based methods for classification of sleep stages based on Stacked Sparse AutoEncoder (SSAE) and Convolutional Neural Network (CNN). The proposed methods performed more efficiently by eliminating the need for conventional feature selection and feature extraction steps respectively. Moreover, although our systems were trained with lower number of samples compared to the similar studies, they were able to achieve state of art accuracy and higher overall sensitivity.

Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Convolutional neural networks have proven to excel at image classification tasks, do to this they have being incorporated into the remote sensing field, initial hurdles in their application like the need for large data sets or heavy computational burden, have being solve with several approaches. In this paper the transfer learning approach is tested for classification of a very high resolution images of a palm oil plantation. This approach uses a pre trained convolutional neural network to extract features from an image, and label them with the aid of machine learning models. The results presented in this study show that the features extracted are a viable option for image classification with the aid of machine learning models. An overall accuracy of 97% in image classification was obtained with the support vector machine model.

Image classification has become a part of our daily routine whether it is classifying between traffic signals or different types of species. However, to differentiate between similar texture and shapes is a difficult task with a naked eye. Latest advancements in the field of computer vision can make this task of image classification easier with deep learning techniques, especially neural networks. However, training neural networks require large datasets, otherwise, it cannot give accurate classification. Inspite all the data availability, there are some subjects which lack enough data. Medical images, rare animals species to name a few examples with relatively less number of information. In our experiment, we have taken those animal species datasets with a minimum number of data and achieved higher classification accuracy. We have examined the various state-of-the-art neural networks like DenseNet and Convolutional Neural Networks that could classify between various animal breeds, and flower species. Furthermore, we compared their results based on accuracy achieved on the test set to determine the most efficient approach. Thus, we could assess which network is most suited for image classification. Moreover, we proposed a two-phase algorithm which differentiates between multiple image dataset through transfer learning via pre-trained Convolutional Neural Network. Initially, images are automatically segmented with the Fully connected network to allow localization of the subject through minimum bounding box around it. Second, we built a robust convolution neural network fine-tuned with a dense network according to our vi image datasets. We also proposed novel steps during the training stage to ensure a robust, accurate and real-time classification. Finally, we have evaluated our method on the well known dog breed dataset, and bird species dataset. The experimental results outclass the earlier methods and achieve an accuracy of 95% to 97% for classifying these datasets.

This research presents a case study at a South African University, involving students who had studied mathematics in a pre-undergraduate Foundation Programme (FP) and who were currently in their first year of study in Information Technology (IT) at the same institution. The study investigated a possible relationship between the teaching approach used in the FP mathematics classroom and the extent of students’ abilities to use important mathematical practices, such as using procedures flexibly; using representation; understanding/explaining concepts; questioning; justifying claims; disagreeing; strategising; and generalising, in an undergraduate IT context. Focus group interviews and task-based interviews were used to answer three related questions: “To what extent are students aware of differences in teaching approaches between FP mathematics and undergraduate study?”; “To what extent do students believe that their experiences of the teaching approaches in the Foundation Programme mathematics class have helped them in undergraduate study in other courses?”; and “In what ways are the mathematical practices taught in the Foundation Programme used in undergraduate study in IT?” A bricolage of learning theories was used as a framework for understanding the possible relationships between teaching approach, development of mathematical practices and learning transfer. The students in the focus groups described the teaching approach used in the FP mathematics classes as studentcentred, whereas many of the undergraduate IT lectures and tutorials were described as teachercentred. The students felt that the approach used in the FP mathematics classroom was beneficial to further study, in that it taught them how to become responsible for their own learning and brought about deep understanding of the mathematical concepts learned in the FP. The task-based interviews showed that all students used mathematical practices to solve IT problems to a greater or lesser extent. The use of these mathematical practices was best understood as being influenced by all past cognitive, social and cultural experiences, and was therefore not a case of “transfer” in the traditional sense of the word. Instead, the use of mathematical practices could be described as an extreme case of “cognitive accommodation” from a cognitive constructivist perspective, or a case of “generality” from a situative perspective. Furthermore, an inter-relationship emerged between student-centred teaching, students’ productive disposition towards mathematics, and the extent of “transfer” of mathematical practices to the IT domain. This interesting relationship warrants further investigation.