Machine learning has enjoyed astounding practical success in a wide range of applications in recent years-practical success that often hurries ahead of our theoretical understanding. The standard framework for machine learning theory assumes full supervision, that is, training data consists of correctly labeled iid examples from the same task that the learned classifier is supposed to be applied to. However, many practical applications successfully make use of the sheer abundance of data that is currently produced. Such data may not be labeled or may be collected from various sources. The focus of this thesis is to provide theoretical analysis of machine learning regimes where the learner is given such (possibly large amounts) of non-perfect training data. In particular, we investigate the benefits and limitations of learning with unlabeled data in semi-supervised learning and active learning as well as benefits and limitations of learning from data that has been generated by a task that is different from the target task (domain adaptation learning). For all three settings, we propose Probabilistic Lipschitzness to model the relatedness between the labels and the underlying domain space, and we discuss our suggested notion by comparing it to other common data assumptions.

Machine learning has enjoyed astounding practical success in a wide range of applications in recent years-practical success that often hurries ahead of our theoretical understanding. The standard framework for machine learning theory assumes full supervision, that is, training data consists of correctly labeled iid examples from the same task that the learned classifier is supposed to be applied to. However, many practical applications successfully make use of the sheer abundance of data that is currently produced. Such data may not be labeled or may be collected from various sources. The focus of this thesis is to provide theoretical analysis of machine learning regimes where the learner is given such (possibly large amounts) of non-perfect training data. In particular, we investigate the benefits and limitations of learning with unlabeled data in semi-supervised learning and active learning as well as benefits and limitations of learning from data that has been generated by a task that is different from the target task (domain adaptation learning). For all three settings, we propose Probabilistic Lipschitzness to model the relatedness between the labels and the underlying domain space, and we discuss our suggested notion by comparing it to other common data assumptions.

The creation of a high-quality optical character recognition system (OCR) requires a large amount of labeled data. Obtaining, or in other words creating, such a quantity of labeled data is a costly process. This thesis focuses on several methods which efficiently use unlabeled data for the training of an OCR neural network. The proposed methods fall into the category of self-training algorithms. The general approach of all proposed methods can be summarized as follows. Firstly, the seed model is trained on a limited amount of labeled data. Then, the seed model in combination with the language model is used for producing pseudo-labels for unlabeled data. Machine-labeled data are then combined with the training data used for the creation of the seed model and they are used again for the creation of the target model. The successfulness of individual methods is measured on the handwritten ICFHR 2014 Bentham dataset. Experiments were conducted on two datasets which represented different degrees of labeled data availability. The best model trained on the smaller dataset achieved 3.70 CER [%], which is a relative improvement of 42 % in comparison with the seed model, and the best model trained on the bigger dataset achieved 1.90 CER [%], which is a relative improvement of 26 % in comparison with the seed model. This thesis shows that the proposed methods can be efficiently used to improve the OCR error rate by means of unlabeled data.

Traditionally design engineers have used the Factor of Safety method for ensuring that designs do not fail in the field. Access to advanced computational tools and resources have made this process obsolete and new methods to introduce higher levels of reliability in an engineering systems are currently being investigated. However, even though high computational resources are available the computational resources required by reliability analysis procedures leave much to be desired. Furthermore, the regression based surrogate modeling techniques fail when there is discontinuity in the design space, caused by failure mechanisms, when the design is required to perform under severe externalities. Hence, in this research we propose efficient Semi-Supervised Learning based surrogate modeling techniques that will enable accurate estimation of a system's response, even under discontinuity. These methods combine the available set of labeled dataset and unlabeled dataset and provide better models than using labeled data alone. Labeled data is expensive to obtain since the responses have to be evaluated whereas unlabeled data is available in plenty, during reliability estimation, since the PDF information of uncertain variables is assumed to be known. This superior performance is gained by combining the efficiency of Probabilistic Neural Networks (PNN) for classification and Expectation-Maximization (EM) algorithm for treating the unlabeled data as labeled data with hidden labels.

<p dir="ltr">The parallel growth of contemporary machine learning (ML) technologies alongside edge/-fog networking has necessitated the development of novel paradigms to effectively manage their intersection. Specifically, the proliferation of edge devices equipped with data generation and ML model training capabilities has given rise to an alternative paradigm called federated learning (FL), moving away from traditional centralized ML common in cloud-based networks. FL involves training ML models directly on edge devices where data are generated.</p><p dir="ltr">A fundamental challenge of FL lies in the extensive heterogeneity inherent to edge/fog networks, which manifests in various forms such as (i) statistical heterogeneity: edge devices have distinct underlying data distributions, (ii) structural heterogeneity: edge devices have diverse physical hardware, (iii) data quality heterogeneity: edge devices have varying ratios of labeled and unlabeled data, and (iv) adversarial compromise: some edge devices may be compromised by adversarial attacks. This dissertation endeavors to capture and model these intricate relationships at the intersection of FL and highly heterogeneous edge/fog networks. To do so, this dissertation will initially develop closed-form expressions for the trade-offs between ML performance and resource cost considerations within edge/fog networks. Subsequently, it optimizes the fundamental processes of FL, encompassing aspects such as batch size control for stochastic gradient descent (SGD) and sampling for global aggregations. This optimization is jointly formulated with networking considerations, which include communication resource consumption and device-to-device (D2D) cooperation.</p><p dir="ltr">In the former half of the dissertation, the emphasis is first on optimizing device sampling for global aggregations in FL, and then on developing a self-sufficient hierarchical meta-learning approach for FL. These methodologies maximize expected ML model performance while addressing common challenges associated with statistical and system heterogeneity. Novel techniques, such as management of D2D data offloading, adaptive CPU clock cycle control, integration of meta-learning, and much more, enable these methodologies. In particular, the proposed hierarchical meta-learning approach enables rapid integration of new devices in large-scale edge/fog networks.</p><p dir="ltr">The latter half of the dissertation directs its ocus towards emerging forms of heterogeneity in FL scenarios, namely (i) heterogeneity in quantity and quality of local labeled and unlabeled data at edge devices and (ii) heterogeneity in terms of adversarially comprised edge devices. To deal with heterogeneous labeled/unlabeled data across edge networks, this dissertation proposes a novel methodology that enables multi-source to multi-target federated domain adaptation. This proposed methodology views edge devices as sources – devices with mostly labeled data that perform ML model training, or targets - devices with mostly unlabeled data that rely on sources’ ML models, and subsequently optimizes the network relationships. In the final chapter, a novel methodology to improve FL robustness is developed in part by viewing adversarial attacks on FL as a form of heterogeneity.</p>