UMD Theses and Dissertations

Permanent URI for this collectionhttp://hdl.handle.net/1903/3

New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a given thesis/dissertation in DRUM.

More information is available at Theses and Dissertations at University of Maryland Libraries.

Browse

Filters

20211

Settings

Author: search.filters.author.Mavridis, ChristosSubject: search.filters.subject.learning

Search Results

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    LEARNING LATENT REPRESENTATIONS AND INTRINSIC LAWS OF COMPLEX SYSTEMS
    (2021) Mavridis, Christos; Baras, John; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    The phenomenal increase of interest towards intelligent autonomous systems observed in recent years is leading the way towards automation augmented with actual machine intelligence. How can a robot reliably react to and operate within its environment? How is this connected to how humans dynamically process sensory information? The answers to these famous problems are not only linked to deeper understanding, but also to the decision-making potential of autonomous systems in general. Towards this direction, it is important to understand and formally ana- lyze the fundamental properties of learning, as a continuous, dynamic, and adaptive process of acquiring new understanding, knowledge, or skills. The contribution of this dissertation is towards two broad and open learning and control problems with applications in image and sound classification, graph partitioning, reinforcement learning, identification and control of multi-agent systems, intelligent transporta- tion, and human-robot interaction. The first problem is connected to the existence of a universal learning archi- tecture in human cognition, which is a widely accepted conjecture supported by established experimental findings from neuroscience. Towards this goal, we study the properties of learning with progressively growing models, and propose the Online Deterministic Annealing (ODA) algorithm that serves as a hierarchical, progressive, interpretable, and knowledge-based learning framework, that can be viewed as an open-box deep learning architecture that requires minimal hyper-parameter tuning. We make use of the mathematical theory and properties of the ODA algorithm to develop efficient and adaptive learning algorithms not only for unsupervised and supervised learning, but also for reinforcement learning, graph partitioning, and detection of leaders in networked systems. Leader detection in networked systems is connected to the second problem we consider: learning the interaction laws of complex collectives, ranging from animal flocks to social networks. This is a time-dependent learning problem with dynamical constraints, where data are often noisy and sparse, and lies beyond the traditional boundaries of machine learning algorithms. We use the ODA algorithm to infer the leadership structure of the networked system. We then adopt an energy-based port-Hamiltonian modeling framework and large-scale optimization techniques to learn the intrinsic structure and interaction laws of the system, which can be used to design defense mechanisms against adversarial UAV swarm attacks. Finally, to study real-life animal flocks and their coordination, we study the interaction laws of networked systems in the macroscopic scale. Due to the fact that real-life observations of the agents of an animal flock are rarely available, we propose a novel learning algorithm based on mathematical principles from mean-field game-theory, to infer the coordination laws of large swarms by observing the evolution of their density over time. This is the first time such a progressive learning approach has been developed and studied in the context of decision-making in such a diverse research area. The insights provided by this work can lead to new developments in machine intelligence based on autonomous, continuously adaptive algorithms that can be used reliably in real-life applications to improve quality of life.