Monocular depth estimation, the task of inferring depth information from a single RGB image, is a fundamental yet challenging problem in computer vision due to its inherently ill-posed nature. This dissertation presents a series of approaches that significantly advance the state-of-the-art in depth estimation.

Currently, the acquisition of accurate cryogenic electron microscopy data deals with problems with complex and time-consuming processes, low signal-to-noise ratio, and missing wedge, leading to a lack of highly accurate imaging data. Such data would be necessary to develop computational methods/visualizations and essential to train deep learning models that are used to solve inverse problems.

With the ever-increasing amount of cyberthreats out there, securing IT and OT infrastructures against these threats has become not only desirable, but fundamental. Network Intrusion Detection Systems (NIDS) are key assets for system protection, providing early alerts of network attacks. An important class of NIDS are those based on ML techniques, around which a substantial amount of research is being done these days. Unfortunately, being ML-based, these NIDS can be targeted by adversarial evasion attacks (AEA), which malicious parties try to exploit to perform network attacks without being detected.

The first part of the dissertation presents a study on the convergence properties of Stein Variational Gradient Descent (SVGD), a sampling algorithm with applications in machine learning. The research delves into the theoretical analysis of SVGD in the population limit, focusing on its behavior under various conditions, including the Talagrand’s inequality T1 and the (L0, L1)−smoothness condition. The study also introduces an improved version of SVGD with importance weights, demonstrating its potential to accelerate convergence and enhance stability.

Deep Learning and generative Artificial Intelligence has grown rapidly during the past few years due to the advancement of computing powers and parallel distributed training algorithms. As a result, it has been a common practice to use hundreds or thousands of machines to train very large Deep Neural Networks.

In this talk, I will present some work in progress on practical optimization methods for deep learning. We will start with a discussion of several empirical techniques that enable training of large-scale models in language and vision tasks, including weight decay, averaging, and schedulers. We will then look at a new approach that we call schedule-free due to its ability to work without a pre-defined time horizon. I will share some details about the theory for these methods, explain why they might be useful in practice and then shed some light on their limitations. This talk will be oriented towards people who already have some knowledge of optimization methods.

We will discuss the solution of eigenvalue problems associated with partial differential equations (PDE)s that can be written in the generalised form Ax = λMx, where the matrices A and/or M may depend on a scalar parameter. Parameter dependent matrices occur frequently when stabilised formulations are used for the numerical approximation of PDEs. With the help of classical numerical examples we will show that the presence of one (or both) parameters can produce unexpected results.

In this talk, I will discuss our progress in advancing the discovery of catalysts for green hydrogen production and carbon dioxide conversion, as well as designing novel metalorganic frameworks for direct air capture.

To improve the data transmission speed of HTTP, HTTP/2 has extended  features based on HTTP/1.1 such as stream multiplexing. Along with its  wide deployment in popular web servers, numerous vulnerabilities are exposed. Denial of service, one of the most popular HTTP/2 vulnerabilities is attributed to the inappropriate implementations of flow control for stream multiplexing.

Most existing AI learning methods can be categorized into supervised, semi-supervised, and unsupervised methods. These approaches rely on defining empirical risks or losses on the provided labeled and/or unlabeled data. Beyond extracting learning signals from labeled/unlabeled training data, in this talk, I will cover a class of methods that I have been developing for over a decade, which can learn beyond the vocabulary that was trained on and can compose or create novel concepts.

Our daily lives are becoming increasingly dependent on smart cyber-physical infrastructures, such as smart homes and cities, smart grid, smart transportation, smart healthcare, smart agriculture, and so on.

Ontologies and Knowledge Graphs are becoming increasingly popular for knowledge representation and reasoning, with a fundamental role in AI and Information Systems.

Networks are nowadays pervasive in Big Data. It is often useful to regroup such data in clusters according to distinctive node features and use a representative element for each cluster, hence generating a novel contracted graph that shrank in size.

The robotics industry has manufactured multiple successful robots that are deployed in various domains and have been playing a significant role in the modern economy.

Despite being small and simple structured in comparison to their victims, virus particles have the potential to harm severly and even kill highly developed species such as humans. To face upcoming virus pandemics, detailed quantitative biophysical un- derstanding of intracellular virus replication mechanisms is crucial. Unveiling the relationship of form and function will allow to determine putative attack points relevant for the systematic development of direct antiviral agents (DAA) and potent vacci- nes. Biophysical investigations of spatio-temporal dynamics of intracellular virus replication so far are rare.

Computational imaging systems are based on the joint design of optics and associated image reconstruction algorithms. Of particular interest in recent years has been the development of end-to-end learned “Deep Optics” systems that use differentiable optical simulation in combination with backpropagation to simultaneously learn optical design and deep network post-processing for applications such as hyperspectral imaging, HDR, or extended depth of field. In this talk I will in particular focus on new developments that expand the design space of such systems from simple DOE optics to compound refractive optics and mixtures of different types of optical components.

Democratizing NLP across numerous languages is a non-trivial task, as it may encounter challenges related to data scarcity, limitations in computational resources, and the intricacies of multilingual and multicultural diversity. The speaker will discuss the efforts and findings in tackling these challenges in this talk. To begin, data scarcity and inconsistency in metadata present common obstacles in low-resource NLP, complicating the understanding of the NLP landscape for low-resource languages.

Remote Direct Memory Access (RDMA) has long been recognized as a powerful technology for high-performance computing and data-intensive applications. In this talk, I will present our experience in deploying intra-region RDMA to support storage workloads in Azure.

The emergence of large language models in text generation has markedly transformed our technological environment, significantly impacting our daily digital interactions.

Datacenters are the cornerstone of our digital lives since they can be viewed as just the other end of our smartphones. From an infrastructure point of view, although they started as a scale-out exercise for commodity off-the-shelf hardware, over the last years we are observing a shift from that paradigm with the emergence of increasingly fast network and storage IO devices, programmable accelerators, and new fast interconnects.

To protect privacy of training data for deep learning models, one line of work proposes to use Differential Privacy (DP). Over recent years, a substantial body of research has emerged, proposing a diverse array of differentially private training algorithms tailored to various deep learning models.

 

Abstract

The talk will give an overview of research at the Department of Computer

We present a theoretical analysis of the Weak Adversarial Networks (WAN) method, recently proposed in [1, 2], as a method for approximating the solution of partial differential equations in high dimensions and tested in the framework of inverse problems. In a very general abstract framework.

Most existing AI learning methods can be categorized into supervised, semi-supervised, and unsupervised methods. These approaches rely on defining empirical risks or losses on the provided labeled and/or unlabeled data.

Analogies between financial markets and critical phenomena have long been observed empirically. So far, no convincing theory has emerged that can explain these empirical observations. Here, we take a step towards such a theory by modeling financial markets as a lattice gas.