Labels are essential for training Deep Neural Networks (DNNs), guiding learning with fundamental ground truth. Label quality directly impacts DNN performance and generalization with accurate labels fostering robust predictions. Noisy labels introduce errors and hinder learning, affecting performance adversely. High-quality labels aid convergence, optimizing DNN training towards accurate data distribution representation. Ensuring label accuracy is vital for DNNs’ effective learning, generalization, and real-world performance. Undoubtedly, ensuring the quality of labels is not only critical but also demanding, often entailing considerable resources in terms of time and cost. As the scale of datasets grows, methods such as crowdsourcing have gained traction to expedite the labeling process. However, this approach comes with its own set of challenges, most notably the inherent susceptibility to errors and inaccuracies. For example, it was observed that the accuracy of AlexNet in classifying CIFAR-10 images plummeted from 77% to a mere 10% when labels were subjected to random flips. This stark drop in accuracy exemplifies the magnitude of influence that corrupted or erroneous labels can exert on the performance of DNNs. Such instances underscore the critical relationship between accurate labels and the efficacy of DNNs in understanding and effectively leveraging data. EnsuringDNNrobustness is vital, involving strategies like noise label identification, filtering, and integrating noise patterns into training for resilientmodels. Architectural and loss function design also combats label-related challenges, enhancing DNN adaptability across applications. This thesis investigates the pivotal role of labels in DNN training and their quality impact onmodel performance. Strategies spanning noise recovery, robust learning frameworks, andmulti-label solutions contribute toDNNresilience against noisy labels, advancing both understanding and practical applications. Chapter 1 of this thesis introduces and explains the crucial elements involved in training DNNs, which include data, DNN models, and expert participation. It highlights the complexity introduced by label noise and sets the stage for the diverse methods designed in subsequent chapters to address these aspects comprehensively.

The digital world is evolving toward representing - and serving the interconnection of - natural persons. Instead of depending on the intrastructure of Big Tech companies and governments, users can cooperate and use their hardware to form public infrastructure. Instead of existing by virtue of a reference in some institution's database, users can interact based on a digital representation of their own choosing. It is no longer sufficient to depend on users to act out of system-imposed altruism. A new digital world is emerging that aims to provide systems that respect the rights of users to control their own digital representation. The complete control over one's own representation and all the data that belongs to it is what we know as Self-Sovereignty.

Solutions for digital Self-Sovereignty are wildly sought after, though their solution space remains woefully underexplored. Numerous global entities, e.g., the European Union, have stated their support for Self-Sovereign systems. However, many old problems of peer-to-peer systems that have gone ignored for decennia resurge with the need for Self-Sovereignty. For example, interconnections in peer-to-peer networks are vulnerable to attacks using fake identities and attackers can manipulate peers by depriving them of data. As most deployed peer-to-peer solutions have very little incentive for disruption by attackers, we have seen very few attacks. However, cryptocurrencies have shown that these attacks do surface when there is sizable monetary gain for attackers. In order to secure our future digital society, we must define and study these systems for Self-Sovereignty.

In this thesis we take the first steps toward defining the systems that can power a Self-Sovereign "Web3" ecosystem. In particular, we explore systems that apply Self-Sovereignty for identity, for public infrastructure, and for the execution of shared code. We describe four prototype mechanisms to form a guide for future work and to derive their general properties. Each mechanism is evaluated as realistically as possible. Thereby, this thesis mostly fulfills an exploratory role to guide the further evolution of our digital world.

With the fast maturing of cloud-based technologies, a variety of cloud services proposed by numerous cloud service vendors to help businesses in a range of scales move to cloud solutions easily. In the cloud computing market, the services are categorized with different service models. Each service model has its own distinct characteristics, advantages, and disadvantages. Serverless or Function as a Service (FaaS) is a newborn service model type and it is leading a trend now.

During the last decade, the proliferation of smartphones, social media and streaming services has provoked an explosion of multimedia data. This large amount of image and video sources combined with more powerful and inexpensive computational capabilities brought by the cloud computing paradigm has facilitated the rapid growth of new machine learning models capable of extracting information faster and more accurately. However, the complexity to develop machine learning models has also grown, involving multiple steps, from the acquisition and preparation of data to the training, evaluation and deployment of models. To alleviate this, the leading database providers have started to integrate the predictive capabilities of machine learning directly into their systems. This new approach is known as in-database machine learning, and it brings new interesting properties such as the exploitation of the inherent relational structure of data or the preservation of its privacy and integrity since the inference occurs directly where the data lives. In this work, we present a cloud-native approach to perform in-database machine learning. We have extended SQLFlow, a bridge between SQL engines and machine learning toolkits to support models trained to solve image recognition tasks over image datasets, which meta-information is persisted on a relational database. Furthermore, we have encapsulated the definition of machine learning models on cloud-native workflows that are able to exploit the GPU resources available in a Kubernetes environment. Our research evaluates the scalability of the proposed system regarding the total execution time and GPU utilization. Besides, we are interested in exploring the design of optimized machine learning query plans, where the goal is to choose the optimal among multiple models that cover a range of specific classes to predict from attending its accuracy and execution cost. For that purpose, we have implemented a model repository containing different model variations and evaluated different strategies to optimize the model selection. Our experiments show that optimizing the model selection will lead to more accurate and faster results, especially when a query covers a high number of classes and the number of models that are able to answer them is limited.

In the traditional software development life cycle, development and operation are divided into different departments. The conflict between departments and, besides, the lack of automation usually leads to low software development efficiency and slow software delivery. Thus, the concept of DevOps is introduced, which combines different departments and automates the process to make software delivery faster and easier. The DevOps toolchain is one important component for adopting DevOps. On the other hand, the adaptation of cloud technology, especially serverless computing makes it tempting for us to investigate what benefits serverless computing brings to the DevOps toolchain. In the first research question, we examine the benefits that AWS serverless platforms bring to DevOps toolchain. To answer this research question, we develop a DevOps toolchain hosted in Amazon Web Services (AWS) and leverage the serverless computing service. In addition, we examine what does each serverless computing service brings to the DevOps toolchain, examine how does the performance of the DevOps toolchain changes with or without using serverless computing service. Our research shows that serverless computing services such as AWS Fargate could reduce the cost, operation effort, and improves performance by enabling parallel execution. Our experiments show that in contrast to a toolchain hosted in a traditional cloud server vs the toolchain that was developed by us using serverless computing service could reduce the total runtime of parallel execution up to 65%. In the second research question, we focus on the integrated toolchain build with AWS DevOps tools from AWS serverless platform. We build a demo integrated DevOps toolchain with AWS DevOps tools and compare the integrated toolchain with the non-integrated toolchain that we built. We find that the integrated toolchain significantly reduces the development time by providing an out-of-box solution for the software team. In addition, the better integration with underlying cloud infrastructure provides more functionality such as global monitoring and blue/green deployment. However, we also find that from the experiment that the performance of the integrated toolchain is lower due to the limitation of resources which also come with a high cost.

With the rapid growth of unstructured data across different mediums, it exposes new challenges for its analysis. To overcome this, data processing pipelines are designed with the help of different tools and technologies for the analysis of data at different stages. One of the applications which we find useful for our company is the creation of knowledge graphs for better representation and understanding of relations in the data. Knowledge graph is a structure of representing information where nodes represent the entities and edges define the relationships among them. The construction of a knowledge graph is a process of extracting meaningful information of entities and relations from unstructured textual data and storing it in a graph database. In this project, we are using Neo4j as a graph database for the efficient storage of data in the form of nodes and relations. To achieve this goal, our first research question proposes the architecture and implementation of a data processing pipeline for the construction of knowledge graphs using unstructured textual data. There are three major stages involved in our pipeline and each component is implemented in a microservice architecture. The first stage starts with the parsing of textual documents in two different formats which are PDF and PPT. In the second stage, we are applying natural language processing techniques for the extraction of meaningful information out of this raw text. In the final stage, key pieces of data are stored into a graph database(Neo4j) for the construction of knowledge graphs. We are running our pipeline on a local machine for evaluating the performance and results of each component. The core aspect of retrieving insights from this unstructured data is achieved with the use of natural language processing. In order to investigate more on this component, our second research question examines the cloud based natural language processing services from three renowned providers which are Amazon, Google and IBM. For choosing a suitable service among them, we evaluate their performance on a common data set of category Marketing from wikipedia. Based on our experimental analysis, IBM stands out among them from the perspective of the quality of output, execution time, features and cost. The adoption of a cloud based service not only leads to a faster development of business solutions but also reduces the engineering effort, its cost and maintenance of our custom implementation only with a little cost per our usage.

The Deal Analytics group of PricewaterhouseCoopers Amsterdam has requested a tool for automatising the business-to-business customer analysis. This analysis was performed manually, which left room for performance improvement. This report discusses how the a product was developed which automates the analyses After two weeks of initial research, a complete system was designed and implemented in the subsequent nine weeks. The tool consists of two distinct parts: a front-end and a back-end. The front-end allows the user to customise the analysis to its own preferences, and communicates with the back-end to efficiently perform the analysis. With the help of user evaluations, the front-end has been designed such that it is usable by any PwC employee within the Deals branch.The back-end uses data analysis techniques and machine learning to analyse customer behaviour. Strong points and growth opportunities of a company are found using techniques such as customer segmentation, regression analysis, and cross-sell analysis. The product has been tested using a variety of techniques to ensure that the software does not crash on unexpected input. The final product is evaluated based on the requirements, design goals and success criteria set at the start of the project and can be considered successful.

The apparition of the Internet was a revolution that allowed to connect people all around the world. It was a disruptive technology that helped to shape the modern society as we know it today. However, to trust a person we interact with but we are not able to see is difficult. In order to trust people in this huge network, different digital identity models have been built to attempt to authenticate its users. Still, as the technology and our use of it evolved, the complexity of digital identity increased and problems started to appear. The two main problems are the aggregation of personal information in the databases of powerful IT companies and the user experience that is more complicated than it should be. Those two problems are linked and are caused by identity models that were not designed in a user-centric approach. Each organization uses its own identity system to authenticate their users, which causes the fragmentation of the user’s digital identity that have to register and create identity for all the organizations he is interacting with. Self-sovereign identity is an emerging identity model that makes use of distributed ledger technologies and cryptography to solve these problems. The goal of this model is to return digital identities in the hand of users by placing them at the center of the model. Unifying all their personal information under one identity would greatly improve their experience and help them to have a better control on how it is used by the organizations they’re interacting with. In our first research question, we first examine if the self-sovereign architecture can provide an identity that can be trusted by people and organizations alike. We found that the trust that we can accord to this model is highly related to our capacity to safely manage the different cryptographic keys that are used to secure wallets where user’s information is stored and to secure the communication channels between the different identity owners. The storage of these keys is the subject of our second research question. To understand better the mechanisms, we implement a prototype using the Hyperledger Indy blockchain to discover how these keys and the credentials associated to the digital identity are stored and managed in this ecosystem. We then use the prototype to authenticate a user and make use of its credentials to authorize or deny the access to specific resources on the system to make the link with identity and access management in our third research question.

Imagine being lost in a desert with a bunch of friends, all of a sudden. Survival will be difficult. You will have mirages, distrust among friends and no means to leave landmarks on the sand. Unable to locate yourself, you will have no means to contact people with maps. The best you can do in such a situation is to stay together in the vicinity of each other and look for food and water. By staying together, you can see more and decrease faulty data; thereby increasing your survival probability. Robots when left to explore the moon encounter the same issues. They do not have a Geo-Positioning System to locate them nor do they have a map. They have faulty sensor readings and might find it difficult to contact a human operator on earth all the time to solve issues on the moon. Since everything looks the same, there are no landmarks to memorise. As they walk around, their battery will also get exhausted. The more we equip the robot outside earth, chances of faults do not decrease, they increase. Therefore, there is a need to make primitive robots capable of autonomous exploration. We prefer sending more than one robot, inspired by the success of the collective strength of insects in harsh environments. This thesis aims at engineering collective behaviour for a group of robots in such resource-less environments like the moon. We expect this collective behaviour to perform searching in time-critical events like earthquake-stricken areas. The thesis is designed to be implemented on legged robots called Zebros. Using communication, they will collectively perform activities such that they appear as one body of tightly coupled autonomous units. We design three distinct algorithms for such missions. Emergent behaviour is expected from the robots running these algorithms. The swarm should collectively choose the best among the possible options without disintegrating into subgroups. (https://www.youtube.com/watch?v=Yf3ToRk7YHY&feature=youtu.be)

With the increased availability and affordability of miniature computing devices,
such as sensors, the era of Internet of Things (IoT) has arrived. Meanwhile, the developments of the cellular mobile technologies and emerge of 5G accelerated the adoption of IoT scenarios within mobile networks. Rapid growth in the number of IoT devices has resulted in greater volumes of data being generated and exchanged between various entities. This highlights the need for efficient data transmission between data producers and consumers. To this goal, Nokia Bell Labs has developed a distributed data dissemination system based on Publish/Subscribe messaging protocol and according to the micro-service architecture. Currently, the system is deployed on virtual machines (VM), where the corresponding compiled Java file (Jar file) of each micro-service is running on a separate VM. This approach for deployment might cause a problem, named ’Dependency Hell’ for the continuous integration and continuous development (CI/CD) workflow. Moreover, one of the system’s limitations is that no service discovery is present. Instead, services are connected through hard-coded IP addresses, defined on their corresponding configuration files. Hence, IPs must be changed manually per deployment according to the infrastructure.
To solve the present limitations, this thesis proposes a deployment based on
Kubernetes. Kubernetes is a container orchestration framework that introduces
several benefits for the deployment process including automation, management,
monitoring and scaling of multi-container packaged applications, such as the current Pub/Sub IoT system. A proof-of-concept solution for the deployment of the system using Kubernetes is presented and its implications on the system’s efficiency and scalability are discussed.

A desirable feature of distributed file systems is to enable clients to mount
external namespaces, allowing them to access physically remote data as if it was
transparently stored in the local cluster. This feature is desirable in a variety of
scenarios, such as in enterprise hybrid-cloud deployments, where some data may be stored remote on the cloud while other in the local file system. This feature introduces questions as how to design, implement, and handle various file system semantics, such as different consistency levels and operational performance between the remote and local storage. For instance, a POXIS-complient block based file system is strongly consistent and does not pose the same semantics as a more relaxed eventual consistent object-based cloud storage. Further, it is interesting to benchmark how such an integrated remote storage solution performs and what potential speedups that might be possible.

To examine such questions and provide the remote storage feature in HopsFS,
a scale-out metadata HDFS distribution, this thesis presents an Proof-of-Concept (PoC) implementation that enables the local file system to be transparently integrated with an external remote storage. The PoC is storage provider agnostic and enables a HopsFS client to on-demand pull in and access remote data from an external storage into the local file system.

Currently, the solution works by mounting a snapshot of a remote storage into
HopsFS, a read-only approach. For write and synchronization work, frameworks
that layers stronger consistency models on top of S3 may be used, such as s3mper.
Performance of the implemented feature has been benchmarked against remote
Amazon S3 buckets, showing results of fast throughput for scanning and persisting metadata, but slower (4 Mb/s) for pulling data into the cluster through the Amazon cloud connector S3A. To improve performance, caching by replication of the remote files onto local disk storage has shown to drastically increase the read speed (up to 100x). The work done in this thesis shows a promising PoC that successfully serves remote blocks into a local file system, with future work including supporting writes, improved caching mechanisms, and reflecting off-band changes between the remote and local storage.

The growing DNA data volumes that originate from novel efficient DNA sequencing methods expose new challenges to computer systems used to process this genomic data. BigData technologies in the Hadoop environment, in particular Apache Spark and the Hadoop Distributed File System (HDFS), are increasingly adapted in state-of-the-art Bioinformatic tools. One application domain of such tools is the Variant Calling Workflow (VCW) that is subject of this work’s research. The application of Spark-based open source tools to execute the different VCW results in a tool chain of separate programs. The programs are executed consecutively and consume the data that is produced by the preceding program as input. This data sharing represents an additional workload as the data needs to be transformed into a file format, written to disk and read from disk again by the next application.
In our first research question we examine whether performance can be increased by improving data sharing. As improving measures we propose (1) the elimination of the single output file generation that is native to most open source Bioinformatics tools and (2) the application of the distributed in-memory file system, Alluxio, as in-memory layer for data sharing between any two consecutive VCW applications. While we achieved in our experiments for (1) an impressive performance boost of 17 % we could not improve performance in our experiments for (2).
In our second research questions we investigate how the data throughput can be improved by changing the execution modes of the VCW. The growing DNA data is mainly represented by a larger quantity of DNA samples. Hence, it is important to optimize the VCW execution for multi-sample input. As part of the second research question we propose different execution modes and show in our experiments that concurrent workflow execution can improve the overall runtime performance in our VCW using 3 input samples of 10GB by 15 % compared to sequential execution.

Graph processing is increasingly popular in a variety of scientific and engineering domains. Consequently, graphs and graph-processing algorithms have become increasingly more diverse. Following the big data trend in every computer-related domain, graphs have also become increasingly larger. Processing graphs is requiring more sophisticated computer systems. Important for this thesis, graph-processing systems now need to combine scalability (a grand challenge in computer science) and raw processing power (an endless race), with efficiency especially in cost and energy requirements (a difficult to define and ensure non-functional property for computer systems). New trade-offs between these elements are offered by two important trends in computer systems. First, distributed systems have grown in popularity and cost-efficiency. Second, GPUs offer an excellent performance-energy ratio and are included in most modern computers. By combining distributed CPU-based systems and non-distributed GPU-enabled systems into distributed heterogeneous systems, large-scale graph processing may become possible and efficient. However, many challenges still exist in the area of graph processing before distributed heterogeneous systems can be well understood. In this thesis, we conduct fundamental and applied research to address three major challenges in three research directions of graph processing: application (understanding new data characteristics and sharing graphs), knowledge (evaluating and comparing the performance of various graphprocessing systems), and design (designing new partitioning policies and entire graphprocessing systems that can use both CPUs and GPUs on multiple machines).