Tesi sul tema "Visual tasks analysis"

This thesis considers the tasks involved in exploratory analysis of temporal graph data, and the visual techniques which are able to support these tasks. There has been an enormous increase in the amount and availability of graph (network) data, and in particular, graph data that is changing over time. Understanding the mechanisms involved in temporal change in a graph is of interest to a wide range of disciplines. While the application domain may differ, many of the underlying questions regarding the properties of the graph and mechanism of change are the same. The research area of temporal graph visualisation seeks to address the challenges involved in visually representing change in a graph over time. While most graph visualisation tools focus on static networks, recent research has been directed toward the development of temporal visualisation systems. By representing data using computer-generated graphical forms, Information Visualisation techniques harness human perceptual capabilities to recognise patterns, spot anomalies and outliers, and find relationships within the data. Interacting with these graphical representations allow individuals to explore large datasets and gain further insightinto the relationships between different aspects of the data. Visual approaches are particularly relevant for Exploratory Data Analysis (EDA), where the person performing the analysis may be unfamiliar with the data set, and their goal is to make new discoveries and gain insight through its exploration. However, designing visual systems for EDA can be difficult, as the tasks which a person may wish to carry out during their analysis are not always known at outset. Identifying and understanding the tasks involved in such a process has given rise to a number of task taxonomies which seek to elucidate the tasks and structure them in a useful way. While task taxonomies for static graph analysis exist, no suitable temporal graph taxonomy has yet been developed. The first part of this thesis focusses on the development of such a taxonomy. Through the extension and instantiation of an existing formal task framework for general EDA, a task taxonomy and a task design space are developed specifically for exploration of temporal graph data. The resultant task framework is evaluated with respect to extant classifications and is shown to address a number of deficiencies in task coverage in existing works. Its usefulness in both the design and evaluation processes is also demonstrated. Much research currently surrounds the development of systems and techniques for visual exploration of temporal graphs, but little is known about how the different types of techniques relate to one another and which tasks they are able to support. The second part of this thesis focusses on the possibilities in this area: a design spaceof the possible visual encodings for temporal graph data is developed, and extant techniques are classified into this space, revealing potential combinations of encodings which have not yet been employed. These may prove interesting opportunities for further research and the development of novel techniques. The third part of this work addresses the need to understand the types of analysis the different visual techniques support, and indeed whether new techniques are required. The techniques which are able to support the different task dimensions are considered. This task-technique mapping reveals that visual exploration of temporalgraph data requires techniques not only from temporal graph visualisation, but also from static graph visualisation and comparison, and temporal visualisation. A number of tasks which are unsupported or less-well supported, which could prove interesting opportunities for future research, are identified. The taxonomies, design spaces, and mappings in this work bring order to the range of potential tasks of interest when exploring temporal graph data and the assortmentof techniques developed to visualise this type of data, and are designed to be of use in both the design and evaluation of temporal graph visualisation systems.

Visual analytics, defined as "the science of analytical reasoning facilitated by interactive visual interfaces," emerged several years ago as a new research field. While it has seen rapid growth for its first five years of existence, the main focus of visual analytics research has been on developing new techniques and systems rather than identifying how people conduct analysis and how visual analytics tools can help the process and the product of sensemaking. The intelligence analysis community in particular has not been fully examined in visual analytics research even though intelligence analysts are one of the major target users for which visual analytics systems are built. The lack of understanding about how analysts work and how they can benefit from visual analytics systems has created a gap between tools being developed and real world practices. This dissertation is motivated by the observation that existing models of sensemaking/intelligence analysis do not adequately characterize the analysis process and that many visual analytics tools do not truly meet user needs and are not being used effectively by intelligence analysts. I argue that visual analytics research needs to adopt successful HCI practices to better support user tasks and add utility to current work practices. As the first step, my research aims (1) to understand work processes and practices of intelligence analysts and (2) to evaluate a visual analytics system in order to identify where and how visual analytics tools can assist. By characterizing the analysis process and identifying leverage points for future visual analytics tools through empirical studies, I suggest a set of design guidelines and implications that can be used for both designing and evaluating future visual analytics systems.

La vision par ordinateur a connu une évolution importante, passant des méthodes traditionnelles aux modèles d'apprentissage profond. L’un des principaux objectifs des tâches de vision par ordinateur est d’émuler la perception humaine. En effet, le processus classique effectué par les modèles d’apprentissage profond dépend entièrement des caractéristiques visuelles, reflétant simplement la manière dont les humains perçoivent visuellement leur environnement. Cependant, pour que les humains comprennent l’environnement qui les entoure, leur raisonnement dépend non seulement de leurs capacités visuelles, mais aussi de leurs connaissances pré-acquises. Combler cette différence entre la perception humaine et celle des machines est essentielle afin de parvenir à un raisonnement similaire à celui des humains. Dans cette thèse, nous proposons de nouvelles approches pour améliorer les performances des modèles d’apprentissage profond en intégrant les systèmes basés sur les connaissances dans les réseaux de neuronaux profonds. L'objectif est d’aider ces réseaux à prendre les bonnes décisions en exploitant à la fois les caractéristiques visuelles et les connaissances pour émuler l’analyse visuelle de l’être humain. Ces méthodologies impliquent deux axes principaux. Premièrement, définir la représentation des connaissances pour incorporer des informations utiles à une tâche spécifique de vision. Deuxièmement, examiner comment intégrer ces connaissances dans les réseaux de neurones pour améliorer leurs performances. La première contribution porte sur l'estimation de la profondeur monoculaire. En effet, les humains sont capables d'estimer leur distance par rapport aux objets perçus, même en n’utilisant qu’un seul œil, et ceci en se basant sur les indices monoculaires. Nous proposons d'intégrer ces indices au sein des réseaux de neurones comme un raisonnement similaire à celui des humains pour l'estimation de la profondeur. À cette fin, nous suggérons d'exploiter un modèle ontologique pour représenter l'environnement comme un ensemble de concepts liés par des relations sémantiques. Les informations sur les indices monoculaires sont extraites grâce à un raisonnement effectué sur l'ontologie proposée et sont transférées dans les réseaux de neurones. Le deuxième travail porte sur la tâche de segmentation panoptique qui vise à identifier toutes les instances d’objets capturées dans une image. Nous proposons une approche qui combine les avantages des réseaux de neurones avec des connaissances sur les relations spatiales entre les objets. Nous avons choisi ce type de connaissances car elles peuvent fournir des indices utiles pour résoudre les ambiguïtés et distinguer entre les instances d'objets similaires. Plus précisément, nous proposons une stratégie d'entraînement qui intègre les connaissances dans le processus d'optimisation des réseaux de neurones. L’approche comprend un processus d'extraction et de représentation des connaissances sur les relations spatiales, qui sont incorporées dans l’entraînement sous forme d'une fonction de perte. Afin de valider l'efficacité des approches proposées, nous avons choisi l'environnement urbain et les véhicules autonomes comme principale cas d’application. Ce domaine est particulièrement intéressant car il s'agit d'un axe de recherche novateur en développement continu, avec des implications significatives pour la sécurité et la mobilité des humains. En conclusion, nous avons étudié diverses approches pour représenter les connaissances et les intégrer aux réseaux de neurones. Ces approches valident que l’utilisation combinée de méthodes basées sur les connaissances et celles basées sur les données conduit de manière constante à des résultats améliorés. Le défi principal réside toujours dans le choix des connaissances pertinentes pour chaque tâche, leur représentation et leur intégration de la manière la plus optimale dans l'architecture des réseaus de neurones profonds
Computer vision has made an important evolution starting from traditional methods to advanced Deep Learning (DL) models. One of the goals of computer vision tasks is to effectively emulate human perception. The classical process of DL models is completely dependent on visual features, which only reflects how humans visually perceive their surroundings. However, for humans to comprehensively understand their environment, their reasoning not only depends on what they see but also on their pre-acquired knowledge. Addressing this gap is essential as achieving human-like reasoning requires a seamless combination of data-driven and knowledge-driven methods. In this thesis, we propose new approaches to improve the performance of DL models by integrating Knowledge-Based Systems (KBS) within Deep Neural Networks (DNNs). The goal is to empower these networks to make informed decisions by leveraging both visual features and knowledge to emulate human-like visual analysis. These methodologies involve two main axes. First, define the representation of KBS to incorporate useful information for a specific computer vision task. Second, investigate how to integrate this knowledge into DNNs to enhance their performance. To do so, we worked on two main contributions. The first work focuses on monocular depth estimation. Considering humans as an example, they can estimate their distance with respect to seen objects, even using just one eye, based on what is called monocular cues. Our contribution involves integrating these monocular cues as human-like reasoning for monocular depth estimation within DNNs. For this purpose, we investigate the possibility of directly integrating geometric and semantic information into the monocular depth estimation process. We suggest using an ontology model in a DL context to represent the environment as a structured set of concepts linked with semantic relationships. Monocular cues information is extracted through reasoning performed on the proposed ontology and is fed together with the RGB image in a multi-stream way into the DNNs. Our approach is validated and evaluated on widespread benchmark datasets. The second work focuses on panoptic segmentation task that aims to identify and analyze all objects captured in an image. More precisely, we propose a new informed deep learning approach that combines the strengths of DNNs with some additional knowledge about spatial relationships between objects. We have chosen spatial relationships knowledge for this task because it can provide useful cues for resolving ambiguities, distinguishing between overlapping or similar object instances, and capturing the holistic structure of the scene. More precisely, we propose a novel training methodology that integrates knowledge directly into the DNNs optimization process. Our approach includes a process for extracting and representing spatial relationships knowledge, which is incorporated into the training using a specially designed loss function. The performance of the proposed method was also evaluated on various challenging datasets. To validate the effectiveness of the proposed approaches for combining KBS and DNNs regarding different methodologies, we have chosen the urban environment and autonomous vehicles as our main use case application. This domain is particularly interesting because it is a challenging and novel field in continuous development, with significant implications for the safety, comfort and mobility of humans. As a conclusion, the proposed approaches validate that the integration of knowledge-driven and data-driven methods consistently leads to improved results. Integration improves the learning process for DNNs and enhances results of computer vision tasks, providing more accurate predictions. The challenge always lies in choosing the relevant knowledge for each task, representing it in the best structure to leverage meaningful information, and integrating it most optimally into the DNN architecture

This Master Thesis, inspired by the relevance that home-offices have assumed as a consequence of the COVID-19 pandemic, originated from the architectural, quantitative, and qualitative analysis of lighting within the author’s student accommodation on KTH Main Campus, in Stockholm (Sweden). As a matter of fact, this background study proved that recommended lighting levels (Illuminance, Contrast Ratio, CCT, CRI, M-EDI, and DER values) were not met, making the space unsuitable and unhealthy to both live and work in. Moreover, evaluating individual lighting preferences, it was demonstrated that flexibility, customizability, and adaptability to daylighting and tasks are fundamental requirements for working from home.Considering the temporary nature of students’ leases, the idea of developing an inexpensive and easily assemblable kit of smart products, aimed at improving lighting in dormitories, came quite naturally.To study the implementation of visual comfort and focus on tasks, literature review, market analysis, quantitative measurements, and hands-on experiments were carried on.Thus, with the gained knowledge and experience, the Home-Office Lighting Kit, an array of commercially available sockets, shades, and smart sources, supporting circadian rhythm and individual preferences, was proposed. Along with it, instructions on set-up and scenarios were provided.Although the overall conclusion was that the market needs products with a better balance between price, sturdiness, and reliability, the lighting conditions within the case study space were substantially improved, making other students interested in doing the same.Eventually, interesting cues for the sector’s development were proposed.

This study assessed the effects of the use of a task analysis on the reliability of visual inspection for naïve participants. Three undergraduate students in a Special Education program took part in the study. Participants chose responses to graphs displaying functional analysis results in a multielement format, without the task analysis, and then with the task analysis. Only one out of three participants showed an increase in reliability with the use of the task analysis. However, further analyses showed that changes in participant reliability for certain functions appeared to be related to the use of the task analysis.

The job of a package delivery driver (PDD) is complex and demanding. These drivers must possess many skills in order to succeed in their work, including physical stamina, appropriate decision-making, positive customer interaction, and most importantly, operational safety. Companies must use significant resources, not only to provide insurance for existing drivers, but also to train new drivers to use their visual attention effectively while driving, and companies have a vested interest in ensuring that the most capable trainees are selected for jobs. Currently, subjective assessments of supervisors or managers are typically used to make these determinations. While these are valuable methods for assessing drivers, an objective measure of how well the driver is using his/her visual attention would both assist evaluators in making judgments, as well as make those judgments more accurate. The purpose of the study described herein was to 1) conduct a task analysis of the driving component of the PDD job responsibilities, and 2) create and test an objective measure that a package delivery company could use to evaluate the performance of its drivers. A detailed task analysis based on numerous observations of drivers in their normal work routines was conducted for this research in order to understand these complex tasks. A framework was created for understanding this system of tasks, which was then used to organize all tasks that drivers were observed to perform into more general, goal-oriented activities. Using this task analysis, incidents were identified that were observed while drivers were behind the wheel. This information demonstrated that breakdowns were occurring within the tasks drivers were performing and that improved methods of training and evaluations may be needed as a result. A construct of visual behavior called Head Down Time (HTD) was then created and tested. An individual HDT is defined as the sum of time of all eye gazes away from the primary display (i.e. windshield) between two distinct eye gazes at the primary display while the vehicle is in motion. HDT was evaluated for its ability to differentiate levels of experience between drivers, its relationship to types of route on which drivers delivered, and its relationship to the driving-related incidents that were observed. HDTs were shown to be differed significantly between drivers of low and high experience, with experienced drivers displaying shorter durations of HDT when compared to inexperienced drivers. HDTs also differed in duration when analyzed by the type of route upon which drivers operated. Commercial and urban routes, while not significantly different with respect to HDT, were shown to have increased HDT durations when compared to rural routes and, in turn, residential routes were found to have significantly longer HDTs than did rural routes and may have significantly shorter durations compared to commercial and urban. Finally, HDTs that were associated with observed driving incidents in terms of chronological proximity were shown to be of significantly longer duration than were HDTs that were not associated with incidents. All tests were conducted using appropriate statistical measures, including t-tests at a level of α = 0.05 for each dataset. Applications of this research include: 1) improvement of PDD training and evaluation methods through use of a detailed task analysis, 2) improvement in how package delivery companies define incidents and train PDD toward the prevention of incidents based on task analysis and observations as to incident frequency, and 3) the further development of HDT as a possible objective measure to supplement the training and evaluation of PDD.
Master of Science

The purpose of this study is to explore how norms and values are described in children´s literature and compare these descriptions with the democratic norms and values in the curriculum.  The theory is inspired by Dewey’s theory about the democratic society. We use, concepts that we have identified in the preschool curriculum that define democratic values. The selection of books that we studied have been chosen from a digital multilingual picture book service called polyglutt which is used in preschools all over Sweden.  The method we used to obtain the result was visual text analysis with a qualitative method approach. Previous research shows a collective view of which values are important in preschool and society. The common theme that we saw in the analysis result was the value of care which was shown in many ways in both text and illustrations in all the books.  The result has shown different ways the characters can represent democratic values through their actions. The books allow children the opportunity to understand and relate to different values. In the discussion the importance of the opportunities that children’s books can contribute to their knowledge of democracy. We have discussed the importance of the preschool teacher’s purpose and knowledge surrounding the literature that is present in the preschool.
Syftet med detta arbete är att beskriva hur värdegrundsfrågor behandlas i barnlitteratur och jämföra dessa beskrivningar med den värdegrund som presenteras i läroplanen. Teorin är inspirerad av Deweys teori om det demokratiska samhället. Vi använder de begrepp om värderingar som finns i förskolans läroplan. Urvalet av böckerna som vi analyserat har valts från en inläsningstjänst som heter polyglutt och som används på förskolor i hela Sverige. Metoden som tillämpades för att få fram resultatet var visuell textanalys med en kvalitativ ansats. Tidigare forskning visar en gemensam syn på vilka värderingar som är viktiga i förskolan samt samhället. Det som var ett övergripande tema i resultatet var värderingen omsorg där vi kunde se omsorg i både bild och text i samtliga böcker på många olika sätt.    Resultatet visar olika sätt som karaktärerna kan förmedla demokratiska värderingar på genom sina handlingar. Böckerna ger barnen möjlighet till att relatera till och förstå de olika värderingarna.    I diskussionsdelen lyfts vikten av barnlitteraturens möjligheter för barns kunskap om demokrati och vikten av förskollärarens syfte med och kunskap kring den barnlitteratur som ska finnas på förskolan.

Group contingencies are efficient and effective behavioral interventions that allow teachers to apply a reinforcement criterion to a large group of students. However, most research on group contingencies has not examined the impact of types of teacher feedback and student choice of teacher feedback incorporated into the use of group contingencies. The current study used a multiple baseline across participants design with an embedded alternating treatments design to explore the use of an interdependent group contingency that compared the effectiveness of incorporating auditory or visual feedback to improve student on-task behavior of three students in public elementary school classrooms. The study also explored whether incorporating student choice into the feedback would enhance the outcomes for student behavior. The results indicated that the interdependent group contingency intervention was successful in increasing the on-task behavior of all three participants. The results also indicated that while both auditory and visual feedback were effective in increasing on-task behavior of all three students, two of the students engaged in slightly higher levels of on-task behavior when auditory feedback was used. When students were given the option to choose which type of feedback would be used, two of the three students favored auditory feedback over visual feedback, and on-task behaviors maintained for all three participants. These results have implications for the use of auditory feedback and choice in the classroom setting as part of a group contingency.

Orientador: Rita de Cássia Ietto Montilha
Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Ciências Médicas
Made available in DSpace on 2018-08-21T17:53:59Z (GMT). No. of bitstreams: 1 Becker_Paula_M.pdf: 1826117 bytes, checksum: bfd083591567eec2a5b180b4fb2ad2e4 (MD5) Previous issue date: 2012
Resumo: Pessoas com deficiência visual deparam-se, na maioria das vezes, com dificuldades que envolvem as esferas social, econômica e funcional, podendo apresentar prejuízo em seu nível de independência e autoestima, comprometendo sua qualidade de vida. O objetivo deste estudo foi identificar níveis de desempenho ocupacional e de qualidade de vida de indivíduos com deficiência visual e posterior análise de inter-relação entre os índices encontrados. Realizou-se um levantamento descritivo e de corte transversal, junto a pessoas com deficiência visual, de idade igual ou superior a 18 anos, inscritas para reabilitação em um serviço universitário, no período de agosto de 2011 a março de 2012. Para a coleta de dados foram utilizados três instrumentos aplicados por entrevista pela própria pesquisadora. O primeiro foi um questionário que permite obter os dados de identificação e perfil sócio-demográfico. O segundo foi o COPM que mensura a auto-percepção do sujeito em relação ao seu desempenho ocupacional e o terceiro foi o SF-36, instrumento que possibilita verificar a auto-percepção do sujeito em relação a sua qualidade de vida. Para análise dos dados foi construído banco de dados e realizado o tratamento estatístico. A amostra foi não probabilística constituída de acordo com os critérios de conveniência e composta de 23 sujeitos, sendo 74,0% com baixa visão, 52,2% do gênero feminino e a média de idade foi de 46,7 anos. A auto-percepção de desempenho ocupacional dos entrevistados foi baixa. Os resultados encontrados mostraram que a auto-percepção de desempenho e aspectos emocionais dos participantes com baixa-visão foram melhores do que os com cegueira. Verificou-se que quanto maior era o tempo de deficiência visual, pior era a avaliação do domínio dor. O domínio vitalidade apresentou relação estatisticamente significativa com os domínios estado geral de saúde, desempenho e satisfação, assim como o domínio saúde mental apresentou relação com estado geral de saúde, dor, desempenho e vitalidade. Os resultados mostram que quanto melhor era o aspecto emocional, maior também a influência positiva refletida nos aspectos físicos, funcionais e sociais dos participantes. O desempenho ocupacional e a qualidade de vida são condições que podem ser trabalhadas em programa de reabilitação para pessoas com deficiência visual por meio de equipe interdisciplinar
Abstract: People with visual disabilities are oftentimes faced with difficulties regarding the social, economic and functional areas, which may impair their level of independence and self-esteem, affecting therefore their quality of life. The purpose of this study was to identify levels of occupational performance and quality of life of individuals with visual disabilities and subsequent analysis of the interrelationship among the indexes found. A descriptive and cross-sectional study was conducted, with a sample group of people with visual disabilities, aged 18 years old and over, which were enrolled on rehabilitation in a university center, from August 2011 to March 2012. For data collection three instruments were applied through interviews by the researcher. The first was a questionnaire used in order to obtain identification data and social-demographic profile. The second was the COPM which measures one's self-perception of occupational performance and the third was the SF-36, an instrument that allows verifying the self-perception of quality of life. For data analysis a database was constructed and statistical analysis was performed. Group sample was classified under a non-probability sampling (convenience) and comprised 23 individuals, 74.0% with low vision, 52.2% were female and the mean age was 46.7 years. Self-perception of occupational performance by the interviewees was low. Results showed that self-perception of performance and emotion aspects of the participants with low vision were better than those with blindness. It was verified that the larger the time of visual impairment, the worse the evaluation of the pain domain. The vitality domain showed a statistically significant relationship with the domains general health, performance and satisfaction, whereas the mental health domain was related to general health, pain, vitality and performance. The results showed that the better the emotional aspect, the greater the positive influence reflected in physical, functional and social aspects of the participants. Occupational performance and quality of life are conditions that may be employed in a rehabilitation program for people with visual disabilities through an interdisciplinary team
Mestrado
Interdisciplinaridade e Reabilitação
Mestra em Saúde, Interdisciplinaridade e Reabilitação

BACKGROUND: Prospective memory difficulties are one of the most common deficits following acquired brain injury. The application of smartphones as a compensatory aid to these difficulties has shown promising results. This study looked to investigate these benefits further. OBJECTIVE: The aims of this study were to investigate whether receipt of reminder prompts through ones smartphone improved completion of pre-planned tasks, in addition to whether it also had secondary implications for participant's wellbeing, confidence, independent functioning, and whether it had any impact on caregiver strain levels. METHOD: This study used an ABAB case series design with mild to moderate acquired brain injury. Task completion rates were monitored across four phases (prompts vs. no prompts). Quantitative questionnaires were administered pre, post and at three months follow up to assess coping with memory difficulties. A qualitative questionnaire explored the perceived impact of the smartphone reminders on everyday functioning, in addition to a 3 month follow up measure assessing attrition rates in smartphone use. RESULTS: Visual inspection analysis suggested greater task completion when reminders were provided. The quantitative questionnaires showed increased use of a Smartphone as reminder device post intervention and at follow up. A basic thematic analysis highlighted a perception that the smartphone system increased task completion, confidence in coping with memory demands, supported emotional wellbeing and reduced dependence on others. As a memory aid it was also less stigmatising and promoted dignity. The three month follow up questionnaire highlighted that all participants continued to use their smartphone as a memory aid. CONCLUSIONS: Use of a smartphone as a memory compensation aid may improve completion of pre-set tasks. Secondary benefits may include increased confidence in coping with memory demands, reduced dependence on others for help, and reduced anxiety or frustration around forgetting.

Schizophrenia is a neurological disorder that is primarily associated with hallucinations, delusions and disordered or confused thinking. Additionally, and less noticeably, the disorder is accompanied by a plethora of cognitive deficits and, as this thesis will address, sensory deficits. Historically, sensory deficits in patients with schizophrenia have been largely discounted. These deficits have instead been attributed to higher-order mechanisms such as impairments to attentional or emotional processing and often localised to the pre-frontal cortex or the limbic system. It is only within the last 50 years (Javitt, 2009) that schizophrenia research has awoken to the possibility of sensory deficits localised to the primary sensory cortices themselves. This thesis will focus on cataloguing potential mechanisms within the visual system which cannot be explained by higher-order processes alone but instead serve as an example of processing deficits in patients with schizophrenia which are likely to be present across the whole brain. Specifically, this thesis will investigate the feasibility of utilising established visual phenomena, through psychophysical and electrophysiological experimentation, within schizophrenia research. This document will first detail our current understanding of sensory deficits in patients with schizophrenia, with particular focus on lower level mechanisms such as gain control. It will also consider how they may be implemented within a clinical setting; this framework will inform all subsequent experiments. Three main experimental chapters will follow titled: "Assessing the functional significance of the Horizontal Effect"; "Correlations between visual oddball responses and schizotypy"; and "Adaptation to Orientation Regularity". Each of these experiments represents an attempt at creating a novel visual paradigm for use within clinical populations which will advance our knowledge of how the visual system performs in patients with schizophrenia. The final chapter will draw upon the conclusions reached from each of the experimental programs and assess future directions for both the current program as well as the field itself.

Texture segmentation is an effortless process in scene analysis, yet its neural mechanisms are not sufficiently understood. A common assumption in most current approaches is that texture segmentation is a vision problem. However, considering that texture is basically a surface property, this assumption can at times be misleading. One interesting possibility is that texture may be more intimately related with touch than with vision. Recent neurophysiological findings showed that receptive fields (RFs) for touch resemble that of vision, albeit with some subtle differences. To leverage on this, here I propose three ways to investigate the tactile receptive fields in the context of texture processing: (1) performance, (2) development, and (3) analysis. For performance, I tested how such distinct properties in tactile receptive fields can affect texture segmentation performance, as compared to that of visual receptive fields. Preliminary results suggest that touch has an advantage over vision in texture segmentation. These results support the idea that texture is fundamentally a tactile (surface) property. The next question is what drives the two types of RFs, visual and tactile, to become different during cortical development? I investigated the possibility that tactile RF and visual RF emerge based on the same cortical learning process, where the only difference is in the input type, natural-scene-like vs. texture-like. The main result is that RFs trained on natural scenes develop RFs resembling visual RFs, while those trained on texture resemble tactile RFs. These results again suggest a tight link between texture and the tactile modality, from a developmental context. To investigate further the functional properties of these RFs in texture processing, the response of tactile RFs and visual RFs were analyzed with manifold learning and with statistical approaches. The results showed that touch-based manifold seems more suitable for texture processing and desirable properties found in visual RF response can carry over to those in the tactile domain. These results are expected to shed new light on the role of tactile perception of texture; help develop more powerful, biologically inspired texture segmentation algorithms; and further clarify the differences and similarities between touch and vision.

A series of fusion techniques are developed and applied to EEG and pupillary recording analysis in a rapid serial visual presentation (RSVP) based image triage task, in order to improve the accuracy of capturing single-trial neural/pupillary signatures (patterns) associated with visual target detection.

The brain response to visual stimuli is not a localized pulse, instead it reflects time-evolving neurophysiological activities distributed selectively in the brain. To capture the evolving spatio-temporal pattern, we divide an extended (``global") EEG data epoch, time-locked to each image stimulus onset, into multiple non-overlapping smaller (``local") temporal windows. While classifiers can be applied on EEG data located in multiple local temporal windows, outputs from local classifiers can be fused to enhance the overall detection performance.

According to the concept of induced/evoked brain rhythms, the EEG response can be decomposed into different oscillatory components and the frequency characteristics for these oscillatory components can be evaluated separately from the temporal characteristics. While the temporal-based analysis achieves fairly accurate detection performance, the frequency-based analysis can improve the overall detection accuracy and robustness further if frequency-based and temporal-based results are fused at the decision level.

Pupillary response provides another modality for a single-trial image triage task. We developed a pupillary response feature construction and selection procedure to extract/select the useful features that help to achieve the best classification performance. The classification results based on both modalities (pupillary and EEG) are further fused at the decision level. Here, the goal is to support increased classification confidence through inherent modality complementarities. The fusion results show significant improvement over classification results using any single modality.

For crucial image triage tasks, multiple image analysts could be asked to evaluate the same set of images to improve the probability of detection and reduce the probability of false positive. We observe significant performance gain by fusing the decisions drawn by multiple analysts.

To develop a practical real-time EEG-based application system, sometimes we have to work with an EEG system that has a limited number of electrodes. We present methods of ranking the channels, identifying a reduced set of EEG channels that can deliver robust classification performance.

Dissertation

With the dynamics of the growing population of native Spanish-speaking students in America’s public schools, two-way dual language immersion programs are increasing in number. A series of observations and interviews in a dual language program by the researcher led to the identification of the use of visuals as a prevalent component integrated in instruction. This dominant instructional strategy led to questions of how visuals might affect student performance. Previous studies had not focused specifically on questions with a visual prompt and questions without a visual prompt. The purpose of this study was to answer several questions focused on differences in student performance on nonvisual and visual questions. The research questions used to achieve this purpose included a comparison of Mathematics TAKS Test scaled scores for students in a two-way dual language immersion program and students enrolled in a monolingual educational program, an analysis comparing these two groups of students on nonvisual and various categories of visual questions, an analysis of the performance of both groups of students based on the comparison of nonvisual questions to subsets of visual questions, and an analysis of an interaction effect and main effects of question type and group based on educational setting. A quasi-experimental design with static-group comparison was used for the purposes of this study. Data were collected from the Texas Education Agency [TEA] for all students who took the Spring 2009 version of the Grade 3 or Grade 4 Mathematics TAKS Test in English or Spanish. Data from a total of 6,904 students were used in the study. Data were analyzed based on descriptive statistics and ANOVAs. A statistically significant difference was found in overall scaled score comparisons between the two groups of students on the Spanish version of the Grade 4 Test. Major findings of the study included a higher performance shown on most subsets of visual questions compared to nonvisual questions for both groups of students involved in this study. Two-way dual language immersion students scored higher than students in a monolingual English program on most subsets of visual questions, but no interaction effect occurred between question type and group.

Yes
The headscarf conceals hair and other external features of a head (such as the ears). It therefore may have implications for the way in which such faces are perceived. Images of faces with hair (H) or alternatively, covered by a headscarf (HS) were used in three experiments. In Experiment 1 participants saw both H and HS faces in a yes/no recognition task in which the external features either remained the same between learning and test (Same) or switched (Switch). Performance was similar for H and HS faces in both the Same and Switch condition, but in the Switch condition it dropped substantially compared to the Same condition. This implies that the mere presence of the headscarf does not reduce performance, rather, the change between the type of external feature (hair or headscarf) causes the drop in performance. In Experiment 2, which used eye-tracking methodology, it was found that almost all fixations were to internal regions, and that there was no difference in the proportion of fixations to external features between the Same and Switch conditions, implying that the headscarf influenced processing by virtue of extrafoveal viewing. In Experiment 3, similarity ratings of the internal features of pairs of HS faces were higher than pairs of H faces, confirming that the internal and external features of a face are perceived as a whole rather than as separate components.
The Educational Charity of the Federation of Ophthalmic and Dispensing Opticians.

University of Technology Sydney. Faculty of Engineering and Information Technology.
We are arriving at the era of big data. The booming of data gives birth to more complicated research objectives, for which it is important to utilize the superior discriminative power brought by explicitly designed feature representations. However, training models based on these features usually requires detailed human annotations, which is being intractable due to the exponential growth of data scale. A possible solution for this problem is to employ a restricted form of training data, while regarding the others as latent variables and performing latent variable inference during the training process. This solution is termed weakly supervised learning, which usually relies on the development of latent variable models. In this dissertation, we propose a novel latent variable model - multinomial latent logistic regression (MLLR), and present a set of applications on utilizing the proposed model on weakly supervised scenarios, which, at the same time, cover multiple practical issues in real-world applications. We first derive the proposed MLLR in Chapter 3, together with theoretical analysis including the concave and convex property, optimization methods, and the comparison with existing latent variable models on structured outputs. Our key discovery is that by performing “maximization” over latent variables and “averaging” over output labels, MLLR is particularly effective when the latent variables have a large set of possible values or no well-defined graphical structure is existed, and when probabilistic analysis is preferred on the output predictions. Based on it, the following three sections will discuss the application of MLLR in a variety of tasks on weakly supervised learning. In Chapter 4, we study the application of MLLR on a novel task of architectural style classification. Due to a unique property of this task that rich inter-class relationships between the recognizing classes make it difficult to describe a building using “hard” assignments of styles, MLLR is believed to be particularly effective due to its ability to produce probabilistic analysis on output predictions in weakly supervised scenarios. Experiments are conducted on a new self-collected dataset, where several interesting discoveries on architectural styles are presented together with the traditional classification task. In Chapter 5, we study the application of MLLR on an extreme case of weakly supervised learning for fine-grained visual categorization. The core challenge here is that the inter-class variance between subordinate categories is very limited, sometimes even lower than the intra-class variance. On the other hand, due to the non-convex objective function, latent variable models including MLLR are usually very sensitive to the initialization. To conquer these problems, we propose a novel multi-task co-localization strategy to perform warm start for MLLR, which in turn takes advantage of the small inter-class variance between subordinate categories by regarding them as related tasks. Experimental results on several benchmarks demonstrate the effectiveness of the proposed method, achieving comparable results with latest methods with stronger supervision. In Chapter 6, we aim to further facilitate and scale weakly supervised learning via a novel knowledge transferring strategy, which introduces detailed domain knowledge from sophisticated methods trained on strongly supervised datasets. The proposed strategy is proved to be applicable in a much larger web scale, especially accounting for the ability of performing noise removal with the help of the transferred domain knowledge. A generalized MLLR is proposed to solve this problem using a combination of strongly and weakly supervised training data.

This thesis includes four main chapters that investigate the use of speeded video as a method of decision-making training. Six different video speeds were tested using a video-based decision-making task, to gain a thorough understanding of the effects of video speed on different levels of expertise in Australian football, particularly, the mechanisms which drive expert performance. The significant difference in performance between elite and sub-elite at the 1.5 speed indicated this to be a potential training speed for elite athletes.