ERC FUNDED PROJECTS

Even after three decades of research on human-computer interaction (HCI), current general-purpose user interfaces (UI) still lack the ability to attribute mental states to their users, i.e. they fail to understand users' intentions and needs and to anticipate their actions. This drastically restricts their interactive capabilities. ANTICIPATE aims to establish the scientific foundations for a new generation of user interfaces that pro-actively adapt to users' future input actions by monitoring their attention and predicting their interaction intentions - thereby significantly improving the naturalness, efficiency, and user experience of the interactions. Realising this vision of anticipatory human-computer interaction requires groundbreaking advances in everyday sensing of user attention from eye and brain activity. We will further pioneer methods to predict entangled user intentions and forecast interactive behaviour with fine temporal granularity during interactions in everyday stationary and mobile settings. Finally, we will develop fundamental interaction paradigms that enable anticipatory UIs to pro-actively adapt to users' attention and intentions in a mindful way. The new capabilities will be demonstrated in four challenging cases: 1) mobile information retrieval, 2) intelligent notification management, 3) Autism diagnosis and monitoring, and 4) computer-based training. Anticipatory human-computer interaction offers a strong complement to existing UI paradigms that only react to user input post-hoc. If successful, ANTICIPATE will deliver the first important building blocks for implementing Theory of Mind in general-purpose UIs. As such, the project has the potential to drastically improve the billions of interactions we perform with computers every day, to trigger a wide range of follow-up research in HCI as well as adjacent areas within and outside computer science, and to act as a key technical enabler for new applications, e.g. in healthcare and education.

1 499 625 €

Start date: 2019-02-01, End date: 2024-01-31

During the last decade, a huge number of earth observation (EO) satellites with optical and Synthetic Aperture Radar sensors onboard have been launched and advances in satellite systems have increased the amount, variety and spatial/spectral resolution of EO data. This has led to massive EO data archives with huge amount of remote sensing (RS) images, from which mining and retrieving useful information are challenging. In view of that, content based image retrieval (CBIR) has attracted great attention in the RS community. However, existing RS CBIR systems have limitations on: i) characterization of high-level semantic content and spectral information present in RS images, and ii) large-scale RS CBIR problems since their search mechanism is time-demanding and not scalable in operational applications. The BigEarth project aims to develop highly innovative feature extraction and content based retrieval methods and tools for RS images, which can significantly improve the state-of-the-art both in the theory and in the tools currently available. To this end, very important scientific and practical problems will be addressed by focusing on the main challenges of Big EO data on RS image characterization, indexing and search from massive archives. In particular, novel methods and tools will be developed, aiming to: 1) characterize and exploit high level semantic content and spectral information present in RS images; 2) extract features directly from the compressed RS images; 3) achieve accurate and scalable RS image indexing and retrieval; and 4) integrate feature representations of different RS image sources into a unified form of feature representation. Moreover, a benchmark archive with high amount of multi-source RS images will be constructed. From an application point of view, the developed methodologies and tools will have a significant impact on many EO data applications, such as accurate and scalable retrieval of: specific man-made structures and burned forest areas.

1 491 479 €

Start date: 2018-04-01, End date: 2023-03-31

"This proposal connects three areas that are considered distant from each other: computational complexity, algebraic geometry, and numerics. In the last decade, it became clear that the fundamental questions of computational complexity (P vs NP) should be studied in algebraic settings, linking them to problems in algebraic geometry. Recent progress on this challenging and very difficult questions led to surprising progress in computational invariant theory, which we want to explore thoroughly. We expect this to lead to solutions of computational problems in invariant theory that currently are considered infeasible. The complexity of Hilbert's null cone (the set of ""singular objects'') appears of paramount importance here. These investigations will also shed new light on the foundational questions of algebraic complexity theory. As an essential new ingredient to achieve this, we will tackle the arising algebraic computational problems by means of approximate numeric computations, taking into account the concept of numerical condition. A related goal of the proposal is to develop a theory of efficient and numerically stable algorithms in algebraic geometry that reflects the properties of structured systems of polynomial equations, possibly with singularities. While there are various heuristics, a satisfactory theory so far only exists for unstructured systems over the complex numbers (recent solution of Smale's 17th problem), which seriously limits its range of applications. In this framework, the quality of numerical algorithms is gauged by a probabilistic analysis that shows small average (or smoothed) running time. One of the main challenges here consists of a probabilistic study of random structured polynomial systems. We will also develop and analyze numerical algorithms for finding or describing the set of real solutions, e.g., in terms of their homology. "

2 297 163 €

Start date: 2019-01-01, End date: 2023-12-31

Over the past 5 years, deep learning has exercised a tremendous and transformational effect on the field of computer vision. However, deep neural networks (DNNs) can only realize their full potential when applied in an end-to-end manner, i.e., when every stage of the processing pipeline is differentiable with respect to the network’s parameters, such that all of those parameters can be optimized together. Such end-to-end learning solutions are still rare for computer vision problems, in particular for dynamic visual scene understanding tasks. Moreover, feed-forward processing, as done in most DNN-based vision approaches, is only a tiny fraction of what the human brain can do. Feedback processes, temporal information processing, and memory mechanisms form an important part of our human scene understanding capabilities. Those mechanisms are currently underexplored in computer vision. The goal of this proposal is to remove this bottleneck and to design end-to-end deep learning approaches that can realize the full potential of DNNs for dynamic visual scene understanding. We will make use of the positive interactions and feedback processes between multiple vision modalities and combine them to work towards a common goal. In addition, we will impart deep learning approaches with a notion of what it means to move through a 3D world by incorporating temporal continuity constraints, as well as by developing novel deep associative and spatial memory mechanisms. The results of this research will enable deep neural networks to reach significantly improved dynamic scene understanding capabilities compared to today’s methods. This will have an immediate positive effect for applications in need for such capabilities, most notably for mobile robotics and intelligent vehicles.

2 000 000 €

Start date: 2018-04-01, End date: 2023-03-31

With the prevalence of information technology in our daily lives, our ability to interact with machines in increasingly simplified and more human-like ways has become paramount. Information is becoming ever more abundant but our access to it is limited not least by technological restraints. Spoken dialogue systems address this issue by providing an intelligent speech interface that facilitates swift, human-like acquisition of information. The advantages of speech interfaces are already evident from the rise of personal assistants such as Siri, Google Assistant, Cortana or Amazon Alexa. In these systems, however, the user is limited to a simple query, and the systems attempt to provide an answer within one or two turns of dialogue. To date, significant parts of these systems are rule-based and do not readily scale to changes in the domain of operation. Furthermore, rule-based systems can be brittle when speech recognition errors occur. The vision of this project is to develop novel dialogue models that provide natural human-computer interaction beyond simple information-seeking dialogues and that continuously evolve as they are being used by exploiting both dialogue and non-dialogue data. Building such robust and intelligent spoken dialogue systems poses serious challenges in artificial intelligence and machine learning. The project will tackle four bottleneck areas that require fundamental research: automated knowledge acquisition, optimisation of complex behaviour, realistic user models and sentiment awareness. Taken together, the proposed solutions have the potential to transform the way we access information in areas as diverse as e-commerce, government, healthcare and education.

1 499 956 €

Start date: 2019-09-01, End date: 2024-08-31

Symmetry is a phenomenon that appears in many different contexts. Algorithmic symmetry detection and exploitation is the concept of finding intrinsic symmetries of a given object and then using these symmetries to our advantage. Application areas of algorithmic symmetry detection and exploitation range from convolutional neural networks in machine learning to computer graphics, chemical data bases and beyond. In contrast to this widespread use, our understanding of the theoretical foundation (namely the graph isomorphism problem) is incomplete and current algorithmic symmetry tools are inadequate for big data applications. Hence, EngageS addresses these key challenges in the field using a systematic approach to the theory and practice of symmetry detection. It thereby also fixes the existing lack of interplay between theory and practice, which is part of the problem. EngageS' main aims are to tackle the classical and descriptive complexity of the graph isomorphism problem and to design the next generation of symmetry detection algorithms. As key ideas to resolve the complexity, EngageS offers three new approaches on how to prove lower bounds and a new method to settle the descriptive complexity. EngageS will also develop practical symmetry detection algorithms for big data, exploiting parallelism and memory hierarchies of modern machines, and will introduce the concept of and a road map to exploiting absence of symmetry. Overall EngageS will establish a comprehensive software library that will serve as a platform for integrated research on the algorithmic treatment of symmetry. In summary, EngageS will develop fast, efficient and accessible symmetry detection tools that will be used to solve complex algorithmic problems in a range of fields including combinatorial algorithms, generation problems, and canonization.

1 999 094 €

Start date: 2019-03-01, End date: 2024-02-29

Probabilistic programs describe recipes on how to infer statistical conclusions about data from a complex mixture of uncertain data and real-world observations. They can represent probabilistic graphical models far beyond the capabilities of Bayesian networks and are expected to have a major impact on machine intelligence. Probabilistic programs are ubiquitous. They steer autonomous robots and self-driving cars, are key to describe security mechanisms, naturally code up randomised algorithms for solving NP-hard problems, and are rapidly encroaching AI. Probabilistic programming aims to make probabilistic modeling and machine learning accessible to the programmer. Probabilistic programs, though typically relatively small in size, are hard to grasp, let alone automatically checkable. Are they doing the right thing? What’s their precision? These questions are notoriously hard — even the most elementary question “does a program halt with probability one?” is “more undecidable” than the halting problem — and can (if at all) be answered with statistical evidence only. Bugs thus easily occur. Hard guarantees are called for. The objective of this project is to enable predictable probabilistic programming. We do so by developing formal verification techniques. Whereas program correctness is pivotal in computer science, the formal verification of probabilistic programs is in its infancy. The project aims to fill this barren landscape by developing program analysis techniques, leveraging model checking, deductive verification, and static analysis. Challenging problems such as checking program equivalence, loop-invariant and parameter synthesis, program repair, program robustness and exact inference using weakest precondition reasoning will be tackled. The techniques will be evaluated in the context of probabilistic graphical models, randomised algorithms, and autonomous robots. FRAPPANT will spearhead formally verifiable probabilistic programming.

2 491 250 €

Start date: 2018-11-01, End date: 2023-10-31

This project is devoted to the analysis of large quantum systems. It is divided in two parts: Part A focuses on the transport properties of interacting lattice models, while Part B concerns the derivation of effective evolution equations for many-body quantum systems. The common theme is the concept of emergent effective theory: simplified models capturing the macroscopic behavior of complex systems. Different systems might share the same effective theory, a phenomenon called universality. A central goal of mathematical physics is to validate these approximations, and to understand the emergence of universality from first principles. Part A: Transport in interacting condensed matter systems. I will study charge and spin transport in 2d systems, such as graphene and topological insulators. These materials attracted enormous interest, because of their remarkable conduction properties. Neglecting many-body interactions, some of these properties can be explained mathematically. In real samples, however, electrons do interact. In order to deal with such complex systems, physicists often rely on uncontrolled expansions, numerical methods, or formal mappings in exactly solvable models. The goal is to rigorously understand the effect of many-body interactions, and to explain the emergence of universality. Part B: Effective dynamics of interacting fermionic systems. I will work on the derivation of effective theories for interacting fermions, in suitable scaling regimes. In the last 18 years, there has been great progress on the rigorous validity of celebrated effective models, e.g. Hartree and Gross-Pitaevskii theory. A lot is known for interacting bosons, for the dynamics and for the equilibrium low energy properties. Much less is known for fermions. The goal is fill the gap by proving the validity of some well-known fermionic effective theories, such as Hartree-Fock and BCS theory in the mean-field scaling, and the quantum Boltzmann equation in the kinetic scaling.

982 625 €

Start date: 2019-02-01, End date: 2024-01-31

Numerical tasks - integration, linear algebra, optimization, the solution of differential equations - form the computational basis of machine intelligence. Currently, human designers pick methods for these tasks from toolboxes. The generic algorithms assembled in such collections tend to be inefficient on any specific task, and can be unsafe when used incorrectly on problems they were not designed for. Research in numerical methods thus carries carries the potential for groundbreaking advancements in the performance and quality of AI. Project PANAMA will develop a framework within which numerical methods can be constructed in an increasingly automated fashion; and within which numerical methods can assess their own suitability, and adapt both model and computations to the task, at runtime. The key tenet is that numerical methods, since they perform tractable computations to estimate a latent quantity, can themselves be interpreted explicitly as active inference agents; thus concepts from machine learning can be translated to the numerical domain. Groundwork for this paradigm - probabilistic numerics - has recently been developed into a rigorous mathematical framework by the PI and others. The proposed research will simultaneously deliver new general theory for the computations of learning machines, and concrete new algorithms for core areas of machine learning. In doing so, Project PANAMA will improve the efficiency and safety of artificial intelligence, addressing scientific, technological and societal challenges affecting Europeans today.

1 450 000 €

Start date: 2018-03-01, End date: 2023-02-28

The systematic investigation of random discrete structures and processes was initiated by Erdős and Rényi in a seminal paper about random graphs in 1960. Since then the study of such objects has become an important topic that has remarkable applications not only in combinatorics, but also in computer science and statistical physics. Random discrete objects have two striking characteristics. First, they often exhibit phase transitions, meaning that only small changes in some typically local control parameter result in dramatic changes of the global structure. Second, several statistics of the models concentrate, that is, although the support of the underlying distribution is large, the random variables usually take values in a small set only. A central topic is the investigation of the fine behaviour, namely the determination of the limiting distribution. Although the current knowledge about random discrete structures is broad, there are many fundamental and long-standing questions with respect to the two key characteristics. In particular, up to a small number of notable exceptions, several well-studied models undoubtedly exhibit phase transitions, but we are not able to understand them from a mathematical viewpoint nor to investigate their fine properties. The goal of the proposed project is to study some prominent open problems whose solution will improve significantly our general understanding of phase transitions and of the fine behaviour in random discrete structures. The objectives include the establishment of phase transitions in random constraint satisfaction problems and the analysis of the limiting distribution of central parameters, like the chromatic number in dense random graphs. All these problems are known to be difficult and fundamental, and the results of this project will open up new avenues for the study of random discrete objects, both sparse and dense.

1 219 462 €

Start date: 2018-04-01, End date: 2023-03-31

Novel technologies like Cloud Computing, Ubiquitous Computing, Big Data, Industry 4.0, and the Internet of Things do not only come with a huge demand for practical and efficient cryptosystems, but also with many novel attack surfaces. The security properties required from cryptographic building blocks for these innovative applications go beyond classical security goals. Modern theoretical cryptography has very successfully developed powerful techniques that enable the design and rigorous formal analysis of cryptosystems in theoretical security models. Now that these techniques are readily available, we have to take the next important step: the evolution of these techniques from idealized theoretical settings to the demands of real-world applications. The REWOCRYPT project will tackle this main research challenge at the intersection of theoretical and real-world cryptography. It will provide a solid foundation for the design and mathematically rigorous security analysis of the next generation of cryptosystems that provably meet real-world security requirements and can safely be used to realize secure communication in trustworthy services and products for a modern interconnected society.

1 498 638 €

Start date: 2019-04-01, End date: 2024-03-31

Despite an improved digital access to scientific publications in the last decades, the fundamental principles of scholarly communication remain unchanged and continue to be largely document-based. The document-oriented workflows in science have reached the limits of adequacy as highlighted by recent discussions on the increasing proliferation of scientific literature, the deficiency of peer-review and the reproducibility crisis. In ScienceGRAPH we aim to develop a novel model for representing, analysing, augmenting and exploiting scholarly communication in a knowledge-based way by expressing and linking scientific contributions and related artefacts through semantically rich, interlinked knowledge graphs. The model is based on deep semantic representation of scientific contributions, their manual, crowd-sourced and automatic augmentation and finally the intuitive exploration and interaction employing question answering on the resulting ScienceGRAPH base. Currently, knowledge graphs are still confined to representing encyclopaedic, factual information. ScienceGRAPH advances the state-of-the-art by enabling to represent complex interdisciplinary scientific information including fine-grained provenance preservation, discourse capture, evolution tracing and concept drift. Also, we will demonstrate that we can synergistically combine automated extraction and augmentation techniques, with large-scale collaboration to reach an unprecedented level of knowledge graph breadth and depth. As a result, we expect a paradigm shift in the methods of academic discourse towards knowledge-based information flows, which facilitate completely new ways of search and exploration. The efficiency and effectiveness of scholarly communication will significant increase, since ambiguities are reduced, reproducibility is facilitated, redundancy is avoided, provenance and contributions can be better traced and the interconnections of research contributions are made more explicit and transparent.

1 996 250 €

Start date: 2019-05-01, End date: 2024-04-30

"The TOROS project targets the challenge of implementing safety-critical cyber-physical systems (CPSs) on commodity multicore processors such that their temporal correctness can be certified in a formal, trustworthy manner. While today it is in principle possible to construct a CPS in a temporally sound way, in practice this rarely happens because, with the current real-time foundations, the prerequisite investments in time, expertise, and resources are prohibitive. This situation is caused in large parts by three fundamental shortcomings in the design of state-of-the-art real-time operating systems (RTOSs) and the applicable timing analyses: (i) current RTOSs expose primarily low-level mechanisms that suffer from accidental unpredictability, i.e., mechanisms that require too much expertise to be used and composed in a temporally sound way; (ii) most analyses rely on idealized worst-case execution-time assumptions that realistically cannot be satisfied on commodity multicore platforms; and (iii) the available real-time theory depends on often complex and tedious proofs, and cannot always be trusted to be sound. As a result, formal timing analysis is rarely relied upon in the certification of CPSs in reality, and instead the use of ad-hoc, unsound ""safety margins"" prevails. The TOROS project seeks to close this gap by moving the RTOS closer to analysis, the analysis closer to reality, and by ensuring that the analysis can be trusted. Specifically, the TOROS project will 1. introduce a radically new, theory-oriented RTOS that by design ensures that the temporal behavior of any workload can be analyzed (even if the application developer is unaware of the relevant theory), 2. develop a matching novel timing analysis that allows for below-worst-case provisioning with analytically sound safety margins that yields meaningful probabilistic response-time guarantees, and 3. mechanize and verify all supporting timing analysis with the Coq proof assistant."

1 499 813 €

Start date: 2019-01-01, End date: 2023-12-31

Pathological communication between different brain regions has been implicated in various neurological disorders. However, the computational tools for assessing such communication from neuroimaging data are not sufficiently developed. The goal of TrueBrainConnect is to establish brain connectivity analysis using non-invasive electrophysiology as a practical and reliable neuroscience tool. To achieve this, we will develop novel signal processing and machine learning techniques that address shortcomings in state-of-the-art reconstruction and localization of neural activity from sensor data, the estimation of genuine neural interactions, the prediction of external (e.g., clinical) variables from estimated neural interactions, and the interpretation of the resulting models. These techniques will be thoroughly validated and then made publicly available. We will use the TrueBrainConnect methodology to characterize the neural bases underlying dementia and Parkinson's disease (PD), two of the most pressing neurological health challenges of our time. In collaboration with clinical experts, we will address practically relevant issues such as how to determine the onset of 'freezing' episodes in PD patients, and how to detect different variants and precursors of dementia. The outcome of TrueBrainConnect will be a versatile methodology allowing researchers, for the first time, to reliably estimate and anatomically localize important types of interactions between different brain structures in humans within known confidence bounds. The proposed clinical applications will improve our understanding of the studied diseases and will lay the foundation for the development of novel diagnostic markers for these diseases.

1 499 875 €

Start date: 2019-01-01, End date: 2023-12-31