CNI College

This thesis contributes to the field of deep latent variable generative models. We provide a comprehensive background on deep generative modeling, covering different model types, neural network parameterization, and learning algorithms. Advancements in latent variable models are proposed by enhancing both model density estimation capabilities and the quality of the learned representations. Based on these contributions, the thesis is structured into two parts. The first part addresses research questions related to enhancing density estimation performance by exploring better probabilistic modeling approaches. First, we focus on a trainable prior distribution and an optimal prior defined by the prior works. Based on this concept, we develop a training framework for continual learning of the variational autoencoders and propose a deep hierarchical VAE which exhibits superior performance and better training stability. We next analyze diffusion models, another class of latent variable models. We study their generative and denoising abilities and propose DAED, a combination of the diffusion model and denoising autoencoder, where these two functions are explicitly decoupled. The second part explores properties of the learned latent representations. First, we show how the hidden data representation learned by the model can be made robust to adversarial attack. Second, we explore the ability of the latent variable model to preserve symmetries of the data and show how these can benefit downstream applications.

Greenhouse gases trap heat near earth's surface, causing warming. Unlike solar warming, which heats the entire atmosphere, greenhouse gases primarily warm the lower atmosphere while cooling the upper layers. Since the industrial revolution, human activities, primarily burning fossil fuels for energy, have significantly increased atmospheric $CO_2$ levels. This increase is the main driver of recent climate change. The global context of rising temperatures necessitates sustainable energy solutions. This thesis reveals critical limitations within current energy management systems, particularly in multi-occupancy buildings. A key research problem is the lack of integrated systems that effectively balance occupant well-being with energy efficiency and privacy. This thesis investigates the development and evaluation of Multimodel Energy Management Systems (MEnMS) for smart buildings, addressing the critical challenges of energy optimization, occupant comfort through personalization, privacy, and sustainability. The research addresses four primary challenges: dynamic spatio-temporal variation in personalized comfort, privacy concerns in IoT-enabled systems, real-time translation of energy consumption to carbon emissions, and the design of a sustainable EnMS adaptable for multiple stakeholders. An IoT-enabled, location-aware smart energy management system is designed and developed, demonstrating significant energy savings and enhanced occupant comfort through the optimization of dynamic lighting configuration. A method combining federated learning with a two-state Markov model is implemented to ensure privacy in personalized EnMS, achieving high accuracy in appliance state prediction while preserving data confidentiality. An integrated MEnMS utilizing IoT architecture is proposed, achieving an average 25% of energy savings through hierarchical control and real-time data analysis for the lighting system. The EnSAF framework, built using the SAF Toolkit, is introduced, providing a structured approach to design sustainability-aware EnMS by addressing diverse stakeholder concerns and utilizing a Decision Map and Software Quality Model. The research contributions include the development of algorithms for occupant localization and appliance preference prediction, the design of an integrated energy management framework, and the creation of a sustainability assessment framework. The implementation of energy management systems in buildings is crucial to reduce energy consumption and carbon emissions. EnMS offers economic and environmental benefits by optimizing energy use. The achievement of energy sustainability requires the collaboration of all stakeholders to build resilient and sustainable communities. The thesis proposes and evaluates several innovative solutions. The thesis concludes by revisiting the research questions, summarizing the findings, and suggesting future directions, including the integration of HVAC systems, the expansion of appliance state prediction algorithms, the analysis of IoT infrastructure overhead, and the refinement of the EnSAF framework for enhanced usability and carbon neutrality.

Over the past decade, we have seen a trend towards centralized digital platforms, which we define as platforms in which the middleman plays a controlling role and typically takes a significant amount of the profit. While the internet is decentralized, Big Tech companies have introduced centralized digital platforms as a means to extract value from society as a business model. Value extraction is a phenomenon that we observe in centralized structures in society. Centralized platforms may have the ambition to control the global market, and in fact want to have a monopoly in that market. Centralized platforms determine the terms & conditions of their platform, which may lead to undesired situations, such as in the case of a content-driven platform where censorship lies in wait. Decentralized Governance to prevent Value Extraction from Society. From a societal perspective, decentralization is a response to design digital ecosystems in which (1) value extraction is prevented, (2) transparency is achieved, (3) decision making in terms of defining, monitoring, and executing are governed decentralized. In a decentralized design approach there is more than one party in both the operating ecosystem (producing the actual economic value), as well as in governing the digital ecosystem. We position this as decentralized governance, in which the rules of monitoring, decision making and execution are decentralized over multiple parties. If decentralized governance should serve as an alternative for the centralized digital ecosystems, governance should not only be defined in terms of parties, rules, legislation, incentive, and decision-making procedures, but more importantly, should also be supported by technology, to cope with the fast-evolving internet-enabling centralized competitors. Blockchain technology might be a possible solution that may support decentralized governance design. If blockchain technology is used to design a decentralized digital ecosystems, equivalent technology should support the corresponding decentralized governance design. DECENT Software Modeling Toolkit to Design Decentralized Governance. We consider blockchain technology as a tool to support decentralized governance, we argue that with respect to decentralized governance, most blockchain technologies are only in their preliminary phase. To become really useful, decentralized digital ecosystems should support on-chain governance that is comprehensible and transparent by design. We consider governance as a topic of design: there is a clear need for software tooling and a modeling language that can contribute in designing decentralized digital ecosystems. Thus, a model- based design is an effective approach to manage and design these complex digital ecosystems. To provide scholars and system designers with an intuitive model-based approach of decentralized governance we propose the DECENT software modeling toolkit. This allows for design, analysis, and evaluation of governance models for digital ecosystems. DECENT has been developed and assessed with industry strength cases in the domain of Peer-to-Peer Energy Trading, Fractional Reserve Banking and Digital Euro.

Large knowledge graphs capture information of a large number of entities and their relations. Among the many relations they capture, class subsumption assertions are usually present and expressed using the rdfs:subClassOf construct. From our examination, publicly available knowledge graphs contain many potentially erroneous cyclic subclass relations, a problem that can be exacerbated when different knowledge graphs are integrated as Linked Open Data. In this paper, we present an automatic approach for resolving such cycles at scale using automated reasoning by encoding the problem of cycle-resolving to a MAXSAT solver. The approach is tested on the LOD-a-lot dataset, and compared against a semi-automatic version of our algorithm. We show how the number of removed triples is a trade-off against the efficiency of the algorithm. The code and the resulting cycle-free class hierarchy of the LOD-a-lot are published at www.submassive.cc.

The emergence of Deep Learning has marked a profound shift in the paradigm of machine learning, a change driven by the numerous breakthroughs it has achieved in recent years. However, as the field evolves and Deep Learning becomes increasingly present in everyday tools and applications, there is an increasing need to address unresolved challenges related to its efficiency and sustainability. This dissertation delves into the role of inductive biases, particularly continuous modeling and symmetry preservation, to address these challenges and enhance the efficiency of Deep Learning. The dissertation is structured in two main parts. The first part investigates continuous modeling as a tool to improve the efficiency of Deep Learning algorithms. Continuous modeling involves the idea of parameterizing neural operations directly in a continuous space. The research presented in this part highlights the substantial benefits of continuous modeling for the (i) computational efficiency –in time and memory–, (ii) the parameter efficiency, and (iii) the complexity of designing neural architectures for new datasets and tasks, coined "design efficiency". In the second half, the focus shifts towards the influence of symmetry preservation on the efficiency of Deep Learning algorithms. Symmetry preservation involves designing neural operations that align with the inherent symmetries of data. The research presented in this part highlights significant gains both in data and parameter efficiency through the use of symmetry preservation. However, it also acknowledges a resulting trade-off of increased computational costs. The dissertation concludes with a thorough critical evaluation of the research findings, openly discussing their limitations and proposing strategies to address them, informed by literature and the author's insights. It ends by identifying promising future research avenues in the exploration of inductive biases for efficiency, and their wider implications for Deep Learning.