L'Internet des Objets (IdO) et les Environnements Intelligents (EI) ont attiré beaucoup d'activités de recherche et développement au cours de la dernière décennie. Pourtant, de nombreuses applications IdO/EI d'aujourd'hui sont encore limitées à l'acquisition et au traitement des données de capteurs et de leur contexte, avec un contrôle, le cas échéant, utilisant soit des solutions de base ou demandant l'intervention humaine, loin du contrôle automatique qui est un facteur essentiel de promouvoir ces technologies. Cette thèse vise à apporter le savoir-faire de la théorie du contrôle et des systèmes réactifs dans le domaine IdO/EI pour arriver à une solution avec une méthode formelle pour l'aspect de contrôle qui fait défaut. Nous proposons l'extension d'un canevas logiciel pour une infrastructure générique et partagée IdO/EI qui offre des interfaces de haut niveau pour réduire l'effort de conception, et qui permet l'auto-configuration et l'adaptation des applications de contrôle sur des propriétés génériques de l'environnement sans intervention humaine en utilisant les connaissances générales sur le domaine qui s'appliquent à chaque instance cible de système IdO/EI. Dans cette infrastructure étendue, les entités physiques individuelles (y compris toutes les "choses", appareils électriques et sous-ensembles de l'espace) peuvent être regroupées comme des entités virtuelles par des propriétés communes afin de fournir un niveau d'abstraction plus élevé pour le contrôle et d'autres applications, ainsi qu'une meilleure adaptation aux changements des configurations au niveau inférieur. Sur le requis d'une solution générique et commun dénominateur partagée par toutes les applications de l'IdO/EI dans un environnement donné, nous proposons pour cette infrastructure, de modéliser les entités cibles supervisées et contrôlées, y compris les entités individuelles et de leurs regroupements, ainsi que les choses et les entités spatiales, par des automates à états finis, pour être en mesure d'appliquer la technique de la synthèse des contrôleur discrets (SCD) aux différents niveaux d'abstraction et de granularité. SCD est une méthode formelle qui construit automatiquement un contrôleur, s'il existe, en assurant les objectifs de contrôle exigés concernant le modèle de comportement du système donné en termes d'automates parallèles synchrones. Les langages de programmation BZR et les outils Sigali existants sont utilisés pour effectuer la SCD et de générer un contrôleur de manière automatique. Les modules logiciels nécessaires sont proposés dans l'implémentation tels que le module de maintenance de relation qui garde une association correcte entre les instances d'entités individuelles et les groupes, et répercute des commandes d'action du contrôle de haut niveau aux actionneurs correspondants. Ce module est destiné à évoluer plus tard vers une solution plus générique comme une base de données graphes comprenant à la fois la base de connaissances générales et relations spécifiques d'instance environnement. La résolution des conflits entre les objectifs de contrôle venant de contrôleurs concurrent est également indispensable en raison des objectifs de l'ouverture de la plateforme. Un simulateur de contexte basé sur Java a été développé pour simuler l'environnement de la maison au sein de plusieurs scénarios proposés pour la validation, tels que le contrôle de la charge électrique et l'adaptation au contexte de l'activité. / The Internet of Things (IoT) and Smart Environments (SE) have attracted a lot of research and development activities during the last decade. Yet many present-day IoT/SE applications are still limited to the acquisition and processing of sensor data and its context, with control, if any, using either basic solutions or requiring human intervention, far away from the automatic control which is an essential factor to promote the technologies. This thesis targets to bring knowhow from control theory and reactive systems to the IoT/SE domain to achieve a solution with a formal method for the missing control aspect. We propose the extension of a framework in order to build a shared generic IoT/SE infrastructure offering high-level interfaces to reduce design effort, and enabling the self-configuration and adaptation of control applications over generic properties of the environment without human interaction by using general knowledge over the domain that applies to each target instance of IoT/SE system. In this extended framework, individual physical entities (including all relevant "things", appliances and subsets of space) may be grouped as virtual entities by shared properties to provide a higher level abstraction for control and other applications and better adaptation to lower level configuration changes. Requiring a generic common denominator solution shared by all IoT/SE applications in a given environment, we propose for this infrastructure, to model by finite state automata the target entities to be monitored and controlled, including both individual entities and their groupings, as well as things and space entities, to be able to apply discrete controller synthesis (DCS) technique over any of these at different levels of abstraction and granularity. DCS is a formal method which constructs automatically a controller, if it exists, guaranteeing the required control objectives regarding to the given system behavior model in terms of synchronous parallel automata. The existing BZR programming language and Sigali tools are employed to perform DCS and generate a controller in an automatic way. Necessary supporting software modules are proposed in the implementation such as the relation maintenance module keeping the correct association between individual entity instances and groups, and dispatching the action orders from the high level control to corresponding actuators. This module would evolve later to a more generic solution such as a graph data base including both the general knowledge base and specific environment instance relations. Conflict resolution between objectives of control coming from concurrent controllers is also indispensable due to the intended openness of the platform. A java based context simulator has been developed to simulate the home environment within several scenarios proposed for the validation, such as electrical load control and activity context adaptation.

Knowledge management has roots in a variety of disciplines, such as philosophy, psychology, social sciences, management sciences and computing. As a result, a wide variety of theories and definitions of knowledge and knowledge management is used in the literature. Irrespective of the theory or definition used, is it recognised that expert knowledge and insight are gained through experience and practice and that it is a key differentiator as an organisational asset. This shift to knowledge as the primary source of value results in the new economy being led by those who manage knowledge effectively. Today’s organisations are creating and leveraging knowledge, data and information at an unprecedented pace – a phenomenon that makes the use of technology not an option, but a necessity. It enables employees to deal with multifaceted environments and problems and make it possible for organisations to expand their knowledge creation capacity. Software tools in knowledge management are a collection of technologies and are not necessarily acquired as a single software solution. Furthermore, these knowledge management software tools have the advantage of using the organisation’s existing information technology infrastructure. Organisations and business decision makers spend a great deal of resources and make significant investments in the latest technology, systems and infrastructure to support knowledge management. It is imperative that these investments are validated properly, made wisely and that the most appropriate technologies and software tools are selected or combined to facilitate knowledge management. The purpose of this interpretive case study is to consider these issues and to focus on an understanding of the key characteristics of a knowledge management system architecture by exploring and describing the nature of knowledge management. Based on the findings of this study, a list of key characteristics that a knowledge management solution must comply with was collated, which expanded the existing knowledge management model towards describing a knowledge management system architecture. / Computing / M.Sc. (Information Systems)

A typical automobile system consists of many Electronic Control Units (ECUs) for the purposes of safety, comfort, and entertainment applications. FlexRay is a high bandwidth protocol for such automotive requirements, which facilitates communication between distributed ECUs. This thesis addresses some of the issues associated with the design and implementation of the FlexRay protocol. The number of ECUs that are used in automobiles has an increasing trend to implement more functionality, which demands more bandwidth. By minimizing bandwidth requirements for servicing a given workload of periodic and sporadic real-time tasks, one can connect more ECUs to the same FlexRay bus. With this motivation, the first part of this thesis proposes new algorithms for minimizing bandwidth usage of ECUs in both the static and the dynamic segments of FlexRay. FlexRay provides a time-triggered static segment for the transmission of time critical periodic messages. The static segment consists of a fixed number of static slots each with a fixed duration. The duration of the static slot and the number of slots are design parameters which need to be fixed so as to meet all the deadline requirements of the workload in the application. The static slot duration can be minimized by packing the signals into message frames, while respecting scheduling constraints. Since message frames also contain overhead information, the duration of message frames can be optimized by proper signal packing, based on the network utilization constraints of individual ECUs. The thesis proposes a novel algorithm for packing of signals into message frames and fixing the static slot duration so that the total duration of the static segment is minimized. The dynamic segment of FlexRay caters to the transmission of event-triggered signals. A novel algorithm has been proposed to obtain the minimum duration of the dynamic segment while meeting the deadline constraints of all sporadic messages in their worst-case arrival instances. We also extend all these algorithms to the case of slot multiplexing scheme provided by FlexRay 3.0. Modern automobiles provide infotainment and in-car telemetry functions, which produce a high volume of soft deadline messages. This makes the problem of analyzing scheduling algorithms for such traffic important. The dynamic segment of the FlexRay cycle is used for transmission of such soft deadline messages. The second part of this thesis addresses the issue of analyzing the quality of performance in servicing of the soft deadline tasks in the FlexRay protocol. Two quality measures, namely, the average delay in servicing of the soft deadline tasks and the fraction of tasks that miss their deadlines, are considered. The generation of different soft deadline messages is modeled as independent Poisson processes. The generated messages are queued in different queues and are serviced according to pre-assigned priorities for different queues as per the FlexRay protocol. By analyzing this multiple queue model under some mild assumptions, upper bounds on the arrival rates for different messages are derived so that all the queues are stable. Analytical expressions are also derived for average delay and for deadline miss ratio. The correctness of these approximate analytical expressions are demonstrated through simulation studies.

O trabalho apresenta um estudo com a finalidade de verificar a aplicabilidade do Modelo do Sistema Viável (VSM Viable System Model) no projeto de uma arquitetura robusta de um Sistema de Informação voltado à área naval nuclear. A ênfase do estudo está em avaliar uma modelagem alternativa para a especificação da arquitetura do Sistema de Informação, incorporando o conjunto de funcionalidades especificadas pelo VSM, com o objetivo de fortalecer esta arquitetura. A estratégia desta pesquisa baseia-se em uma revisão bibliográfica relacionada ao VSM, aos Sistemas de Informação e a sua arquitetura, e às influências destes elementos na sobrevivência das organizações diante das mudanças constantes no ambiente. É apresentado um estudo de caso onde são mostrados os elementos teóricos do VSM e da arquitetura de SI aplicados na elaboração da arquitetura de um SI. O Sistema escolhido para esta aplicação é o Sistema de informação de incidentes nucleares (SIN) nas instalações onde são feitas pesquisas e desenvolvimento de tecnologia nuclear a ser aplicada na propulsão de um submarino. / The present work presents a study aiming to verify the applicability of the Viable System Model (VSM) in a robust architecture for an Information System targeting nuclear naval area. The emphasis of the study is in assessing an alternative modeling for the Information System (IS) architecture specification, incorporating a set of functionalities defined by the VSM, with the purpose of strengthening this architecture. The strategy of this research is based on a bibliographic revision on VSM, Information System and its architecture, and the influence of those elements in the survival of the organizations in a ever changing environment. It is presented one case study where it is showed the theoretical elements of the VSM and IS architecture applied on the development of a IS architecture. The selected system for this application is the IS for nuclear incidents (SIN) on the installations dedicated to research and development on nuclear technology applied to submarine propulsion system.

La sélection des nouveaux porte-greffes du genre Prunus a pour principal objectif d’utiliser lavariabilité génétique existant au sein des différentes espèces de Prunus afin de créer un matérielvégétal innovant, performant au plan agronomique et résistant à différentes contraintes biotiques etabiotiques. L’exploitation de la variabilité génétique s'appuie sur le phénotypage des différentsindividus présents dans les collections de ressources génétiques pour les caractères recherchés. Celapermet de sélectionner des génotypes spécifiques pour améliorer un trait donné.Dans cette étude constituée de deux parties, nous avons exploré la diversité génétique au seindu genre Prunus pour ce qui concerne deux traits importants : l’aptitude au bouturage et les réponsesde certains porte-greffes à la contrainte hydrique.Dans la première partie l’aptitude au bouturage ligneux a été explorée dans une collectiongénétique de 222 génotypes. Les résultats obtenus montrent une variabilité considérable entre les sousgenresAmygdalus et Prunophora et également une variabilité interspécifique au sein de chaque sousgenre.La réussite au bouturage est nettement améliorée chez les hybrides interspécifiques dont un desparents appartient à l’espèce P cerasifera.Dans la deuxième partie nous avons étudié les réponses à la contrainte hydrique chez neufgénotypes couramment utilisés comme porte-greffes. Nous avons comparé trois régimes hydriquesdifférents : témoins, stressés et recouvrés. L’humidité du substrat est maintenue à la capacité au champtout au long de l’expérience pour les plants témoins, par contre l’arrosage a été arrêté pendant 14 jourspour les plants stressés, puis il a été repris pendant 10 jours pour les plants recouvrés. Durant cetteexpérience, nous avons effectué des mesures morphologiques et physiologiques sur la partie aérienneainsi que des excavations à la fin de chaque phase pour examiner les modifications au niveau dusystème racinaire. Sur la partie aérienne, la contrainte hydrique a provoqué une diminutionsignificative de la photosynthèse nette, de la transpiration totale, de la conductance stomatique, et del’expansion des feuilles.Le système racinaire a répondu à cette contrainte par plusieurs modifications. D’abord, le ratio racines/pousses a augmenté pour 4 génotypes (GF305, GF677, Montclar et Myrobolan1254). Ensuite, lesdifférents traits de l’architecture racinaire ont montré des modifications sous l’effet de la contraintehydrique : la longueur de la zone apicale non ramifiée (LZANR) qui traduit l’élongation racinaire,ainsi que les diamètres apicaux des racines ont diminué chez tous les génotypes. Par conséquent lesracines se sont affinées et ont réduit leur croissance en longueur. Produire des racines plus finesaugmente la surface de contact entre les racines et le sol et améliore la capacité d’absorption. Unediminution de la distance inter-ramification a été observée chez les plants stressés. Cette modificationpourrait s’expliquer par le fait que les plantes produisent plus de racines latérales en profondeur oùl’eau est souvent plus disponible. De plus, les racines latérales produites étaient également plus fines.Au plan qualitatif, les génotypes ont eu des réponses semblables, mais l’intensité de la réponse a variéselon les génotypes. / In breeding programs of Prunus rootstocks the aim is to use the existing genetic variabilitywithin Prunus species in order to create new rootstock genotypes with excellent agronomic traits, andimproved resistance to biotic and abiotic stresses.Exploitation of the genetic variability is based on the evaluation of phenotypic variation amongindividuals within genetic collections for desirable traits. This make possible to select specificgenotypes to improve a given trait.This study consists of two parts; we explored the genetic diversity within the genus Prunusregarding two important characteristics: rooting ability of hardwood cuttings and responses of somerootstocks to water stress.In the first part, rooting ability of hardwood cuttings was evaluated among 222 genotypespreserved in genetic collection. The results show considerable variability among the sub genusAmygdalus, Prunophora, and also an interspecific variability within each of them. Rooting ability byhardwood cutting was significantly improved in interspecific hybrids if one parent belongs to Pcerasifera species.In the second part of this study we studied the responses of nine genotypes, commonly usedrootstocks, to water stress. We applied three treatments: control, water stress and recovering.Soil moisture was maintained at field capacity through all stages of the experiment for the controlplants. On the contrary we stopped watering during 14 days for the stressed plants, and then we rewateredthe recovered plants for 10 days.During this experience, we performed morphological and physiological measurements on the aboveground parts of plants and we excavated plants at the end of each phase to observe root systemmodifications.Aboveground parts of plant responded to water stress by a significant decrease in net photosynthesis,total transpiration, stomatal conductance and leaf expansion.Root system responded to water stress by several modifications:Four genotypes (GF305, Montclar, GF677 and Myrobolan) showed a significant increase in root toshoot ratio under drought conditions. We also detected morphological modifications on the differenttraits of root architecture in response to water stress.The length of the apical unbranched zone LAUZ and the apical diameter were decreased forall genotypes, consequently, roots became finer and reduced their rate of elongation. Fine rootsenhance the surface of contact between roots and soil which in turn improve the acquisition of waterunder drought condition.The inter-branch distance also responded, and it tended to decrease under the water stress treatment.The decrease in inter-branch distance can be explained by a production of more lateral roots in deeplayers, where water was more available, moreover these new laterals roots were also finer.Qualitatively, a common response to water stress was observed on the different traits of the rootsystem architecture, but we showed a genotypic effect determining the level of the response.

Abstract At the heart of the smart space vision is the idea that devices interoperate with each other autonomously to assist people in their everyday activities. In order to make this vision a reality, it is important to achieve semantic-level interoperability between devices. The goal of this dissertation is to enable Semantic Web technology-based interoperability in smart spaces. There are many challenges that need to be solved before this goal can be achieved. In this dissertation, the focus has been on the following four challenges: The first challenge is that the Semantic Web technologies have neither been designed for sharing real-time data nor large packets of data such as video and audio files. This makes it challenging to apply them in smart spaces, where it is typical that devices produce and consume this type of data. The second challenge is the verbose syntax and encoding formats of Semantic Web technologies that make it difficult to utilise them in resource-constrained devices and networks. The third challenge is the heterogeneity of smart space communication technologies that makes it difficult to achieve interoperability even at the connectivity level. The fourth challenge is to provide users with simple means to interact with and configure smart spaces where device interoperability is based on Semantic Web technologies. Even though autonomous operation of devices is a core idea in smart spaces, this is still important in order to achieve successful end-user adoption. The main result of this dissertation is a semantic interoperability framework, which consists of following individual contributions: 1) a semantic-level interoperability architecture for smart spaces, 2) a knowledge sharing protocol for resource-constrained devices and networks, and 3) an approach to configuring Semantic Web-based smart spaces. The architecture, protocol and smart space configuration approach are evaluated with several reference implementations of the framework components and proof-of-concept smart spaces that are also key contributions of this dissertation. / Tiivistelmä Älytilavision ydinajatuksena on, että erilaiset laitteet tuottavat yhteistyössä ihmisten elämää helpottavia palveluita. Vision toteutumisen kannalta on tärkeää saavuttaa semanttisen tason yhteentoimivuus laitteiden välillä. Tämän väitöskirjan tavoitteena on mahdollistaa semanttisen webin teknologioihin pohjautuva yhteentoimivuus älytilan laitteiden välillä. Monenlaisia haasteita täytyy ratkaista, ennen kuin tämä tavoite voidaan saavuttaa. Tässä työssä keskityttiin seuraaviin neljään haasteeseen: Ensimmäinen haaste on, että semanttisen webin teknologioita ei ole suunniteltu reaaliaikaiseen kommunikaatioon, eivätkä ne sovellu isojen tiedostojen jakamiseen. Tämän vuoksi on haasteellista hyödyntää niitä älytiloissa, joissa laitteet tyypillisesti jakavat tällaista tietoa. Toinen haaste on, että semanttisen webin teknologiat perustuvat syntakseihin ja koodausformaatteihin, jotka tuottavat laitteiden kannalta tarpeettoman pitkiä viestejä. Tämä tekee niiden hyödyntämisestä hankalaa resurssirajoittuneissa laitteissa ja verkoissa. Kolmas haaste on, että älytiloissa hyödynnetään hyvin erilaisia kommunikaatioteknologioita, minkä vuoksi jopa tiedonsiirto laitteiden välillä on haasteellista. Neljäs haaste on tarjota loppukäyttäjälle helppoja menetelmiä sekä vuorovaikutukseen semanttiseen webiin pohjautuvien älytilojen kanssa että tällaisen älytilan muokkaamiseen käyttäjän tarpeiden mukaiseksi. Vaikka laitteiden itsenäinen toiminta onkin älytilojen perusajatuksia, tämä on kuitenkin tärkeää teknologian hyväksymisen ja käyttöönoton kannalta. Väitöskirjan päätulos on laitteiden semanttisen yhteentoimivuuden viitekehys, joka koostuu seuraavista itsenäisistä kontribuutioista: 1) semanttisen tason yhteentoimivuusarkkitehtuuri älytiloille, 2) tiedonjakoprotokolla resurssirajoittuneille laitteille ja verkoille sekä 3) menetelmä semanttiseen webiin pohjautuvien älytilojen konfigurointiin. Näiden kontribuutioiden evaluointi suoritettiin erilaisten järjestelmäkomponenttien referenssitoteutuksilla ja prototyyppiälytiloilla, jotka kuuluvat myös väitöskirjan keskeisiin kontribuutioihin.

Abstract The contemporary information age is equipped with rich and affordable telecommunication services. In the future, people have even more flexibility when true wireless Internet and real-time multimedia are provided seamlessly over heterogeneous wireless networks. Optimally combining the capacity and services of the current and emerging networks requires a holistic view of mobility, resource and service management. This thesis contributes to the research and development of these hybrid systems with three main contributions. Firstly, a system architecture for vertical handoff in location-aware heterogeneous wireless networks is proposed. The proposed architecture enables the mobile node to prepare for approaching vertical handoffs and wake-up a hotspot interface. The needed communication procedures are discussed, and inter-related issues of mobility and geolocation information are considered in proportion to usability, advantages and limitations. Secondly, a framework for the analysis of vertical handoff algorithm sensitivity to various mobility parameters including velocity, handoff delay and dwell time is introduced. Handoff smoothing with a dwell-timer is analyzed as one potential scheme for optimizing vertical handoff locally. It is compared to a power based algorithm to find out its sensitivity to the changes in effective data rates, velocity of the terminal and the amount of handoff delay. The analysis focuses on the transition region, having case studies on both moving-in and moving-out scenarios. An optimal value for dwell-timer is found through simulations, showing a performance gain over power based algorithm as a function of mean throughput. The analysis is extended also to a multiple network scenario. Thirdly, experimental results on the behaviour of protocols used in wireless IP networks are presented. Prototype systems demonstrate results of using Mobile IP with a fuzzy logic algorithm for vertical handoff in a heterogeneous network environment and the role of IPv6 when using a voice application in a wireless LAN environment. Latest contributions include developing plug-and-play middleware functionalities for Symbian mobile devices, extending the use of the earlier results to state-of-the-art mobile devices.

<p>Deep Neural Networks (DNNs) have greatly advanced the state-of-the-art in a wide range of machine learning tasks involving image, video, speech and text analytics, and are deployed in numerous widely-used products and services. Improvements in the capabilities of hardware platforms such as Graphics Processing Units (GPUs) and specialized accelerators have been instrumental in enabling these advances as they have allowed more complex and accurate networks to be trained and deployed. However, the enormous computational and memory demands of DNNs continue to increase with growing data size and network complexity, posing a continuing challenge to computing system designers. For instance, state-of-the-art image recognition DNNs require hundreds of millions of parameters and hundreds of billions of multiply-accumulate operations while state-of-the-art language models require hundreds of billions of parameters and several trillion operations to process a single input instance. Another major obstacle in the adoption of DNNs, despite their impressive accuracies on a range of datasets, has been their lack of robustness. Specifically, recent efforts have demonstrated that small, carefully-introduced input perturbations can force a DNN to behave in unexpected and erroneous ways, which can have to severe consequences in several safety-critical DNN applications like healthcare and autonomous vehicles. In this dissertation, we explore approximate computing as an avenue to improve the speed and energy efficiency of DNNs, as well as their robustness to input perturbations.</p> <p> </p> <p>Approximate computing involves executing selected computations of an application in an approximate manner, while generating favorable trade-offs between computational efficiency and output quality. The intrinsic error resilience of machine learning applications makes them excellent candidates for approximate computing, allowing us to achieve execution time and energy reductions with minimal effect on the quality of outputs. This dissertation performs a comprehensive analysis of different approximate computing techniques for improving the execution efficiency of DNNs. Complementary to generic approximation techniques like quantization, it identifies approximation opportunities based on the specific characteristics of three popular classes of networks - Feed-forward Neural Networks (FFNNs), Recurrent Neural Networks (RNNs) and Spiking Neural Networks (SNNs), which vary considerably in their network structure and computational patterns.</p> <p> </p> <p>First, in the context of feed-forward neural networks, we identify sparsity, or the presence of zero values in the data structures (activations, weights, gradients and errors), to be a major source of redundancy and therefore, an easy target for approximations. We develop lightweight micro-architectural and instruction set extensions to a general-purpose processor core that enable it to dynamically detect zero values when they are loaded and skip future instructions that are rendered redundant by them. Next, we explore LSTMs (the most widely used class of RNNs), which map sequences from an input space to an output space. We propose hardware-agnostic approximations that dynamically skip redundant symbols in the input sequence and discard redundant elements in the state vector to achieve execution time benefits. Following that, we consider SNNs, which are an emerging class of neural networks that represent and process information in the form of sequences of binary spikes. Observing that spike-triggered updates along synaptic connections are the dominant operation in SNNs, we propose hardware and software techniques to identify connections that can be minimally impact the output quality and deactivate them dynamically, skipping any associated updates.</p> <p> </p> <p>The dissertation also delves into the efficacy of combining multiple approximate computing techniques to improve the execution efficiency of DNNs. In particular, we focus on the combination of quantization, which reduces the precision of DNN data-structures, and pruning, which introduces sparsity in them. We observe that the ability of pruning to reduce the memory demands of quantized DNNs decreases with precision as the overhead of storing non-zero locations alongside the values starts to dominate in different sparse encoding schemes. We analyze this overhead and the overall compression of three different sparse formats across a range of sparsity and precision values and propose a hybrid compression scheme that identifies that optimal sparse format for a pruned low-precision DNN.</p> <p> </p> <p>Along with improved execution efficiency of DNNs, the dissertation explores an additional advantage of approximate computing in the form of improved robustness. We propose ensembles of quantized DNN models with different numerical precisions as a new approach to increase robustness against adversarial attacks. It is based on the observation that quantized neural networks often demonstrate much higher robustness to adversarial attacks than full precision networks, but at the cost of a substantial loss in accuracy on the original (unperturbed) inputs. We overcome this limitation to achieve the best of both worlds, i.e., the higher unperturbed accuracies of the full precision models combined with the higher robustness of the low precision models, by composing them in an ensemble.</p> <p> </p> <p><br></p><p>In summary, this dissertation establishes approximate computing as a promising direction to improve the performance, energy efficiency and robustness of neural networks.</p>

Current airspace regulations require the remote pilot-in-command of an unmanned aircraft systems (UAS) to maintain visual line of sight with the vehicle for situational awareness. The future of UAS will not have these constraints as technology improves and regulations are changed. An operational model for the future of UAS is proposed where a remote operator will monitor remote vehicles with the capability to intervene if needed. One challenge facing this future operational concept is the ability for a flight data system to effectively communicate flight status to the remote operator. A system architecture has been developed to facilitate the implementation of such a flight data system. Utilizing the system architecture framework, a Phase I prototype was designed and built for two vehicles in the Autonomous Flight Laboratory (AFL) at Cal Poly. The project will continue to build on the success of Phase I, culminating in a fully functional command and control system for remote UAS operational testing.

As software systems become increasingly larger and more complex, the need to make them easily maintained increases, as large systems are expected to last for many years. It has been estimated that system maintenance is a large part of many IT-departments’ software develop­ment costs. In order to design a complex system to be maintainable it is necessary to introduce structure, often as models in the form of a system architecture and a system design. As development of complex large-scale systems progresses over time, the models may need to be reconstructed. Perhaps because development may have diverted from the initial plan, or because changes had to be made during implementation. This thesis presents a reconstructed documentation of a complex large-scale system, as well as suggestions for how to improve the existing design based on identified needs and insufficiencies. The work was performed primarily using a qualitative manual code review approach of the source code, and the proposal was generated iteratively. The proposed design was evaluated and it was concluded that it does address the needs and insufficiencies, and that it can be realistically implemented.