About us

Consortium

Team

AI Flanders

Contact

The research program is organized around four Grand Challenges:

Research

Many industrial processes and systems in society impose complex preconditions for making decisions. This research line, subtitled 'Making Data Science Hybrid, Automated, Trusted and Actionable', focuses on making automatic analyzes of available data, formulating existing expertise and generating new knowledge through machine learning. And all of this taking into account the requirements in terms of security, ethics and privacy.

ch1-structure-update.png

storage/airesearch/ch1-wp1.jpg

WP 1 is concerned with all steps starting from the ingestion of the data to the generation of labelled data and high-level features. It will contribute partially automated techniques for data ingestion and wrangling, data quality assessment and enhancement, semantic interpretation, automatic feature generation and active learning.&nbsp;

WP 1 AI-assisted Data Acquisition & Preprocessing

wp2-structure.png

WP 2 focuses on using knowledge in the data science process, and as such it will contribute knowledge representation languages that extend the common ground between learning and reasoning, new algorithms for learning and reasoning as well flexible and interactive ways for humans to supervise and steer learning and reasoning processes.

WP 2 Hybrid Models for Learning and Reasoning

ch1-wp3-update-structure.png

WP 3 will contribute with AI techniques to (partially) automate the data exploration processes, and to better assist data analysts in obtaining interesting and actionable insights from data. It will focus on AI-assisted interactive techniques, on formalizing &lsquo;insights&rsquo; in intelligible and actionable ways, on exploiting background knowledge to ensure the (subjective) interestingness of such discovered insights, on realistic high-dimensional and structured datasets, and on scalable methods.

WP 3 AI-Assisted Data-Exploration

ch1-wp4-structure.png

WP 4 is concerned with the automation of machine learning and as such focuses on the model construction step starting from the data and knowledge produced in the earlier WPs. It will focus on three steps: problem analysis, the actual model construction step and performance tuning while keeping the human in the loop.

WP 4 Automation in Machine Learning

ch1-wp5-update-structure.png

WP 5 focuses on technical AI-oriented solutions for obtaining trustworthy AI systems. It wants to build trust, ethical and privacy aspects directly into the core of AI methods. It focuses on three techniques: explainability, ethics (pen-testing) and preserving the privacy of data in a federated manner.&nbsp;

WP 5 Trustworthy AI

ch1-wp6-update-structure.png

WP 6 focuses on decision support, that is, on the adoption of AI technologies *after* algorithm development, benchmarking, and training of machine learning systems. Decision support systems have an integrative role vis &agrave; vis technologies, and a matchmaking and educative role vis &agrave; vis users and applications. Thus WP 6 contributes to the system, user, process and organizational level of decision support.

WP 6 Decision Support Systems

WP 7 focuses on the following Use Cases in the area of health: Single Cell Technologies, Multiple Sclerosis, Epilepsy, Medical Image Processing, Length of Stay in Hospitals.

WP 7 Use cases in Health

WP 8 focuses on the following Use Cases for industry and society: Production Process Optimization, Low Voltage Energy Distribution Grid, Employment Services, Prognostic and health monitoring for ageing wind turbines.&nbsp;

WP 8 Use Cases in Industry, Energy, for Government & Citizens

Work packages

Management team, <a href="https://www.esat.kuleuven.be/stadius/" target="_blank" rel="noopener">KU Leuven - ESAT/STADIUS</a>

Bart de Moor

Management team, <a href="https://idlab.ugent.be/" target="_blank" rel="noopener">UGent - IDLab</a>

Piet Demeester

Ann Ackaert

Oscar Mauricio Agudelo

WP1 Lead: AI-assisted Data Acquisition and Pre-Processing, <a href="https://dtai.cs.kuleuven.be/" target="_blank" rel="noopener">KU Leuven - CS/DTAI</a>

Luc De Raedt

WP2 Lead: Integrated learning and reasoning, <a href="https://dtai.cs.kuleuven.be/" target="_blank" rel="noopener">KU Leuven - CS/DTAI</a>

Hendrik Blockeel

WP3 Lead: AI-Assisted Data Exploration,<a href="https://idlab.ugent.be/" target="_blank" rel="noopener"> UGent - IDLab</a>

Tijl De Bie

WP4 Lead: Automation in machine learning, <a href="https://idlab.ugent.be/" target="_blank" rel="noopener">UGent - IDLab</a>

Tom Dhaene

WP5 Lead: Trustworthy AI, <a href="https://www.esat.kuleuven.be/stadius/" target="_blank" rel="noopener">KU Leuven - ESAT/STADIUS</a>

Yves Moreau

WP6 Lead: Decision Support Systems, <a href="https://www.esat.kuleuven.be/psi/" target="_blank" rel="noopener">KU Leuven- ESAT/PSI</a>

Matthew Blaschko

WP7 Lead: Use Cases Healthcare, <a href="https://www.dambi.ugent.be/" target="_blank" rel="noopener">UGent - VIB - DAMBI</a>

Yvan Saeys

WP8 Lead: Use Cases Industry, <a href="https://www.esat.kuleuven.be/stadius/" target="_blank" rel="noopener">KU Leuven - ESAT/STADIUS</a>

Bart De Moor

Multiple research groups collaborate on this research domain. This table mentions the contact person and his/her affiliation.

contacts

Making data science hybrid, automated,trusted and actionable

Unlocking the value of data in a trusted and automated manner, supporting complex decision making and providing new insights that will empower individuals and society in generating major advances in healthcare, education, industry 4.0, energy systems and more.

AI-Driven data science

storage/airesearch/research-challenge-aidrivendata-plain-zonder-wit.png

AI-Driven Data Science

Powerful smartphones, cars and robots can take over tasks from the cloud (edge ​​computing), leading to faster decisions, lower energy consumption and better privacy protection. This opens new possibilities for AI applications based on intelligent systems and components with low power, often on batteries. The research will lead to application-oriented cases far ahead of the state of the art for distributed and hierarchical AI systems, advanced signal processing, and learning algorithms for extracting actionable information from the edge.

gc2.png

Management team, <a href="https://www.imec-int.com/en" target="_blank" rel="noopener">imec</a>

Rudy Lauwereins

Management team &amp; WP1 Lead: Use Cases, <a href="https://www.imec-int.com/en" target="_blank" rel="noopener">imec</a>

Axel Nackaerts

WP2 Lead: Inter-device Algorithms, <a href="https://idlab.ugent.be/" target="_blank" rel="noopener">UGent - IDLab</a>

Bart Dhoedt

WP3 Lead: Intra-Device Algorithms, <a href="https://ipi.ugent.be/" target="_blank" rel="noopener">UGent - IPI</a>

Wilfried Philips

WP4 Lead: Software Tool Suites, Software to Hardware Mapping, <a href="https://www.imec-int.com/en" target="_blank" rel="noopener">imec</a>

Peter Debacker

WP5 Lead: Extreme Edge Hardware, <a href="https://www.esat.kuleuven.be/micas/" target="_blank" rel="noopener">KU Leuven - MICAS</a>

Marian Verhelst

Contacts

Real-time and power-efficient AI in the edge

Improving edge device environments through the co-optimisation between power efficient Al processors and advanced machine learning tasks with as purpose to increase the real-time performance, reliable low-latency communication, power-efficient processing and data security.

AI in the edge

storage/airesearch/research-challenge-edge-plain-zonder-wit.png

AI in the Edge

The research in this grand challenge is focused on multi-agent systems. In multi-agent systems, it is fundamental that each agent has its own goals and intentions and none of the agents knows the entire system, nor has direct control over the other agents. Multi-agent systems can be found in many domains, such as trading systems, network routing systems, intelligent traffic systems, smart grids, autonomous vehicles, internet of things, collaborating robots and machines, and many more. The research focuses on the adaptivity, robustness, manageability of multi-agent systems and how to formulate guarantees on their proper functioning.

storage/airesearch/gc3-overview.jpg

WP1 covers 4 use cases:
In the domain of <a href="/en/research/application-domains/industry" data-entry-id="28" target="_blank" rel="noopener" data-entry-id="28">industry</a>:&nbsp; Smart Control of Cooperating Robots / Control of Collaborating Robots Smart Control of Machine Fleet (bar linkages setup and wind turbine setup) / Control of a Machine Park: Similar Machines Learning from each Other
In the domain of <a href="/en/research/application-domains/energy" data-entry-id="29" target="_blank" rel="noopener" data-entry-id="29">energy</a>: &nbsp; Self-Sustainable Smart Grids&nbsp; Verbeterde opbrengst van windmolenparken
In the domain of <a href="/en/research/application-domains/government-citizens"[entry:airesearch:27:/research/application-domains/government-citizens]data-entry-id="27">government &amp; citizens</a>: &nbsp; Exchanging Data and Knowledge in a Healthcare Setting&nbsp; &nbsp; These use cases allow us to validate the newly developed architectures and algorithms in real-world scenarios, and act as a bridge between the researchers of the grand challenge and external stakeholders such as industry and government.&nbsp;

WP 1 Use Cases

To address the increasing demand for flexible robotization in several applications as manufacturing, warehousing, infrastructure, agriculture, rescue and disaster response, etc; the task complexity or flexibility is too high for a single agent to accomplish or the task is inherently distributed in space. Under these circumstances, building several resource-bounded agents or robots is much better than having a single complex robot. Besides their increased robustness through redundancy, one of the main advantages is that multiple agents can solve problems faster using parallelism. Traditional centralized control approaches cannot be used with these systems due to, e.g. centralized computational issues or issues with centralized data collection, actuation and communication and scale constraints. WP2 combines in a multilayer and centralized/decentralized control architecture model-based and model-free methods with human advice to obtain a team of cooperating agents and humans that are able to perform tasks safely from the start and improve their collective performance over time. As a result, the agents do not have a global view of the problem, while, depending on the degree of interaction that exists between the local subsystems, the agents or robots need to coordinate e.g., through strategic communication or via incentives.

WP 2 Multi-Agent Control Systems

WP3 studies aspects of hybrid multi-agent systems, which include human agents as well as digital agents. The presence of humans brings important extra challenges, which the ongoing COVID-19 pandemic has made more obvious than ever. The available measures such as social distancing or wearing masks are all heavily dependent on human compliance. Yet, very often our personal goals are not necessarily aligned with the common good.&nbsp; &nbsp; The goals of this WP are threefold. A first task proposes new techniques for modelling individual and collective behaviour and investigates how this behaviour needs to be integrated into agent systems to make successful hybrid systems. A second task aims to investigate how much trust plays a role in using hybrid MAS and which incentives need to be provided to ensure long-term relations between human participants and MAS. At the moment knowledge of hybrid MAS behaviour is underutilized in industry. Finally, a third task is focusing on how outcome of collective behaviour can be influenced, either through incentives or through exchange of information (which itself will be influenced by how much agents are willing to share their information and the reciprocity of the exchange).

WP 3 Hybrid Multi-Agent Systems for Collective Action

The focus of WP4 is to investigate how to exchange privacy-sensitive data and knowledge in a multi-agent environment, where information requests need to be efficiently matched with data or information sources. When one party is willing to exchange some of their data with another party through an AI system, the parties must be able to communicate with the system. How can they specify to the AI what data to share, and what to keep private, without having to manually indicate everything explicitly, which is impractical? How can a party communicate in what structure or shape they would like to receive the data? How can data be exchanged, transformed, and queried automatically?

WP 4 Distributed Data Intelligence

Work Packages

Management team, <a href="https://ai.vub.ac.be/?utm_source=airesearchflanders.be&amp;referrer-analytics=1" target="_blank" rel="noopener">VUB AI Lab</a>

 Ann Nowé

Management team &amp; WP1 Lead: Use Cases, <a href="https://ai.vub.ac.be/?utm_source=airesearchflanders.be&amp;referrer-analytics=1" target="_blank" rel="noopener">VUB AI Lab</a>

Leander Schietgat

WP2 Lead: Multi-Agent Control Systems, <a href="http://mech.vub.ac.be/" target="_blank" rel="noopener">VUB Robotics &amp; Multibody Mechanics</a>

Bram Vanderborght

WP3 Hybrid Multi-Agent Systems for Collective Action, <a href="https://ai.vub.ac.be/?utm_source=airesearchflanders.be&amp;referrer-analytics=1" target="_blank" rel="noopener">VUB AI Lab</a>

Bart de Boer

WP4 Lead: Distributed Data Intelligence, <a href="https://www.uhasselt.be/UH/Onderzoeksgroepen/DetOndgr.html?oid=86" target="_blank" rel="noopener">UHasselt</a>

Jan Van den Bussche

AI systems that interact autonomously with other decision-making entities&nbsp;

This challenge addresses the fact that in many real-life situations, different autonomous decision-making entities are not standalone, but interact with each other.&nbsp;This requires flexible coordination mechanisms for these autonomous decision-making entities, allowing to adapt to their environments, to interact flawlessly with humans, and to exchange privacy-sensitive data, in order to solve collaborative tasks in a highly connected and rapidly changing world. &nbsp;

Multi-agent Collaborative AI 

storage/airesearch/research-challenge-collaborative-plain-zonder-wit.png

This research line focuses on the design of AI systems to communicate and interact with people as naturally as possible, including through natural language and image recognition. The AI ​​system must be able to build layered human-like reasoning by observing and understanding the complex environment. This allows computers to independently identify and solve problems. Humans can work with their intelligence and physical abilities in harmony with complementary machines through intuitive and social interaction. This goal is still a long way off. But in many practical applications, such as recognizing patterns and generalizing individual tasks, AI already makes a very interesting contribution.

challenge4structure.jpg

WP1, WP2, WP3 and WP4 are work packages that feature both perception and cognition: starting from multimodal sensors, clear semantics are derived, which are used in human-like learning algorithms, with a strong personalization and interaction aspect. The results of these methodological work packages are integrated in WP5: use cases and demonstrators.
Following the objectives set forth above, this challenge can be split in two main parts, which are both addressed in two methodological work packages. The overall work package structure is illustrated in Figure above.
<ul>
<li>How can we get a detailed understanding of the environment at hand? This is tackled in work packages WP 1 and WP 2</li>
<li>How can we respond and react to tasks in a way that is similar to a human? This is tackled in WP 3 and WP 4.</li>
</ul>

WP 1 'Audio-visual perception &amp; multimodal representations' has a dual goal. On the one hand, it will develop new vision and audio analysis techniques to achieve visual and auditive understanding. On the other hand, it focuses on learning a clear and structured representation of the system&rsquo;s surroundings.&nbsp;
WP Lead: Tinne Tuytelaars, ESAT, PSI, &nbsp;KULeuven
- KULeuven ESAT, PSI (<a href="https://www.esat.kuleuven.be/psi" target="_blank" rel="noopener">https://www.esat.kuleuven.be/psi</a>)

WP 1 Audio-visual perception & multimodal representations

WP2 'Deep learning-based conversational agents' focuses on natural language processing and has the goal of creating a deep learning based conversational agent which can incorporate background knowledge, common sense and support interaction. This contextual knowledge can be extracted from other sensors such as vision and audio, studied in WP1.&nbsp;
WP Lead: Walter Daelemans, CLIPS, UAntwerpen
Research Groups:
<ul>
<li><a href="https://www.uantwerpen.be/en/research-groups/clips/" target="_blank" rel="noopener">UAntwerpen CLIPS</a></li>
<li><a href="https://www.lt3.ugent.be/" target="_blank" rel="noopener">UGhent LT3</a></li>
<li><a href="https://ugentt2k.github.io/" target="_blank" rel="noopener">UGhent IDLab</a> (<a href="https://ugentt2k.github.io/" target="_blank" rel="noopener">Text-to-Knowledge Group (T2K) </a></li>
<li><a href="https://liir.cs.kuleuven.be/" target="_blank" rel="noopener">KU Leuven LIIR</a></li>
<li><a href="https://ehai.ai.vub.ac.be/" target="_blank" rel="noopener">VUB AI lab</a></li>
<li><a href="https://clips.github.io/ADATaLKS/researchvideos/" target="_blank" rel="noopener"> In the following videos our PhD students pitch their current research (2021) on Conversational Agents</a></li>
</ul>

WP 2 Deep learning-based conversational agents

WP3 'Interaction, personalization and recommendation' focuses on achieving a natural and intuitive interaction with an AI system and giving the AI system a first &ldquo;direction&rdquo;. Out of a wide variety of choices (i.e., big decisions) an AI system can take, what is the best one that tailors to the human it interacts with? New recommendation techniques which are explainable are key to achieve this.&nbsp; 
WP Lead: Bart Goethals, ADREM, UAntwerpen
Research Groups: - <a href="https://www.uantwerpen.be/en/research-groups/adrem-data-lab/" target="_blank" rel="noopener">UAntwerpen ADREM</a> - <a href="https://www.ugent.be/mict/en" target="_blank" rel="noopener">UGent MICT</a> - <a href="https://www.ugent.be/ea/idlab/en" target="_blank" rel="noopener">UGent IDLab</a> - <a href="https://www.uhasselt.be/edm" target="_blank" rel="noopener">UHasselt EDM</a>&nbsp; - <a href="http://www.etrovub.be/" target="_blank" rel="noopener">VUB ETRO&nbsp;</a> - <a href="https://www.waves.intec.ugent.be/" target="_blank" rel="noopener">UGent WAVES</a>&nbsp;

WP 3 Interaction, personalization and recommendation

WP4 'Coginitive architectures &amp; human-like learning' has the goal of developing new machine learning methods and solutions, which bring the reasoning closer to how a human learns and reasons about the world. This entails fast adaptation to unseen tasks, detecting and learning preferences, etc.&nbsp; WP Lead: Geraint Wiggins, VUB AI Lab&nbsp;
Research Groups:
- <a href="https://ai.vub.ac.be/computational-creativity/" target="_blank" rel="noopener">VUB AI Lab</a> - <a href="https://www.uantwerpen.be/en/research-groups/idlab/" target="_blank" rel="noopener">UAntwerpen IDLab&nbsp;</a> - <a href="https://we.vub.ac.be/nl/dima-digital-mathematics" target="_blank" rel="noopener">VUB DiMa</a> - <a href="https://mobi.research.vub.be/en/home" target="_blank" rel="noopener">VUB MOBI</a> - <a href="https://www.waves.intec.ugent.be/" target="_blank" rel="noopener">UGent WAVES&nbsp;&nbsp;</a>

WP 4 Coginitive architectures & human-like learning

WP 5 Use Cases

Management team, <a href="https://www.imec-int.com/en" target="_blank" rel="noopener">imec</a>, <a href="https://www.uantwerpen.be/en/research-groups/idlab/" target="_blank" rel="noopener">UAntwerpen - IDLab,</a>&nbsp;<a href="mailto:erik.mannens@uantwerpen.be" target="_blank" rel="noopener">e-mail</a>

Erik Mannens

Management team &amp; WP5 Lead: Use Cases, <a href="https://www.uantwerpen.be/en/research-groups/idlab/" target="_blank" rel="noopener">UAntwerpen - IDLab</a>

Tom De Schepper

WP1 Lead: Audio-visual Perception and Multimodal Representations, <a href="https://ai.kuleuven.be/" target="_blank" rel="noopener">KU Leuven</a>

Tinne Tuytelaars

WP2 Lead: Deep Learning-based Conversational Agents, <a href="https://www.uantwerpen.be/en/research-groups/clips/" target="_blank" rel="noopener">UAntwerpen - CLiPS</a>

Walter Daelemans

WP3 Lead: Interaction, Personalization and Recommendation, <a href="https://www.uantwerpen.be/en/research-groups/adrem-data-lab/" target="_blank" rel="noopener">UAntwerpen - ADREM</a>

Bart Goethals

WP4 Lead: Cognitive Architectures &amp; Human-like learning, <a href="https://ai.vub.ac.be/?utm_source=airesearchflanders.be&amp;referrer-analytics=1" target="_blank" rel="noopener">VUB</a>

Geraint Wiggins

AI-systems that communicate and collaborate seamlessly with humans

Towards more natural, interactive, personalized, and human-inspired AI systems. Seamless interaction between humans and AI in Multi-modal perception, Multi-modal instruction, Personalized interaction and responses, Complex control: navigation, reasoning, etc.

Research Challenges

AI-Driven Data Science

AI in the Edge

Multi-agent Collaborative AI

Human-like AI