9th of May 2022
Campus Lágymányos, Northern Building, Harmónia room, -1.85.
Tentative Program
12.00 | 12.10 | Opening, Welcome | János Botzheim, Udo Bub Head of AI Dept, Director of Industry-Academy Cooperation Institute |
12.10 | 12.50 | Keynote lecture: Rehabilitation Robots that Cooperate and Motivate | Robert Riener Full professor, Sensory-Motor Systems Lab, ETH Zurich and SCI Center, University Hospital Balgrist University of Zurich |
12.50 | 13.10 | Humane AI | George Kampis Senior researcher, DFKI |
13.10 | 13.30 | Semantic Matching and Human-Centered AI: A Product-Oriented SWOT Analysis | András Lőrincz Tudományos főmunkatárs, ELTE |
13.30 | 13.50 | How Amazon became an AI-powered company | Attila Papp Lead Cloud Solution Architect @ adesso Hungary Software |
13.50 | 14.30 | Állófogadás, szendvicsebéd | |
14.30 | 14.50 | AI in the MEMS based inertial sensor production | Itilekha Podder European Institute of Innovation & Technology (EIT) Doctoral Student Eötvös Loránd University (ELTE) , Budapest Robert Bosch Kft., Budapest Hungary |
14.50 | 15.10 | Robust Machine Learning | Márk Jelasity Szegedi Tudományegyetem, Természettudományi és Informatikai Kar, Informatikai Intézet, Számitógépes Algoritmusok és Mesterséges Intelligencia Tanszék |
15.10 | 15.30 | Application of AI in self-driving function development and in the product development | Zoltán Kárász Senior AI Expert at Robert Bosch |
15.30 | 15.50 | On the Structural Analysis of Sparse Neural Networks | László Gulyás Department of Artificial Intelligence, Faculty of Informatics, Eötvös Loránd University, Hungary |
15.50 | 16.10 | Embedding dynamic networks in vector spaces on-the-fly | András Benczúr Head of Informatics Laboratory at the Institute for Computer Science and Control, scientific coordinator of the Hungarian Artificial Intelligence National Laboratory |
16.10 | 16.30 | Spiking neural networks and their applications | János Botzheim Department of Artificial Intelligence, Faculty of Informatics, Eötvös Loránd University, Hungary, Head of Department |
Demos
In connection with the opening of the Department of Artificial Intelligence, an interactive exhibition of the research and development results of the department can also be viewed. This is possible before the start of the conference from 11:30 or during the lunch break. Location: 7.103A.
Northern building, room 7.100-7.103:
- Data exploration & anomaly detection – VR experience
- Autonomous driving
- Piano playing robot
Poster session:
- Bacterial programming and it’s applications
- Learning to swarm; machine learning for collective intelligence
- AI in the MEMS-based sensor production
Northern building, room 7.92-7.94:
- Tracking: birds & rats
- DeepRehab: robot-assisted physical rehabilitation
- Semantic mapping
- Neuromorphic moon lander
- Personality trait estimation
Northern building, Community Area (front of the elevator of the Danube side):
- ELTE-Bosch collaboration at the department
- Play a Kahoot game about the secrets of AI and charge yourself with a „Sportszelet”
Entrance of the Globe Hall (Outside the Northern building):
Digital cameras and LiDAR devices in Autonomous Driving
- Bosch’s test car: Tesla showcase
- ELTE’s test car: Skoda and Dodgem car showcase
Program details
Keynote lecture
Robert Riener
Full professor, Sensory-Motor Systems Lab, ETH Zurich and SCI Center, University Hospital Balgrist University of Zurich
Rehabilitation Robots that Cooperate and Motivate
Robots can support the restoration of upper and lower limb movement functions. They promote neuroplastic effects when used as therapy devices after neurological injuries such as spinal cord injury or stroke. This talk will highlight new patient-cooperative controllers, intention detection strategies and virtual reality technologies that improve the rehabilitation of arm movements and gait. Advanced machine learning approaches have been applied for psychological state estimation from psycho-physiological recordings to control patient engagement during robot-assisted gait training. In another project, machine learning technologies were applied to multimodal sensor data sets to predict movement intention in upper extremities reaching and grasping tasks to assist human arm movements with the arm connected to a therapy robot. In the latest activities deep neural networks were applied to detect snoring and use the information for a postural intervention of a robotic bed during sleep.
George Kampis
Senior researcher, DFKI
Humane AI
The talk mainly deals with a topic currently in the trade, and reflected in the title. My excuse for this sweepingly general topic is that (in DFKI, Germany) we coordinate an EU project on it, that comprises 53 leading European laboratories. I detail aspects of the issue and a possible approach to it. Details include social credits systems and the criticism thereof, explainable AI, or the Alphie project at ELTE. Our approach in the above project is based on the notion of micro-projects (MPs). An MP is a partner-initiated response to a semi-open call, adequate for AI. MPs have the advantage that they combine the top-down objectives of the project with the dynamically changing, bottom-up initiatives of the partners and their external collaborators. The purpose of the theme is the emerging EU legal regulation of AI, the „EU AI Act“, which we also dissuss briefly.
András Lőrincz
Tudományos főmunkatárs, ELTE
Semantic Matching and Human-Centered AI: A Product-Oriented SWOT Analysis
The R&D results of the Department of Artificial Intelligence open up innovative opportunities and require a SWOT analysis of potential products. The results start from the basic sciences (structured sparse coding), cover different areas of artificial intelligence (emotion and personality state estimation, hand, head, body pose estimation, tracking, information fusion, natural language processing and semantic mapping) and go up to software developments (deep neural networks and Composite AI). Application examples include “continuous healthcare” and “security of home and public spaces”, among others. Product-oriented integration and industrial-grade software development is what could be the next step. Based on the technology components, it can be said that the SWOT “weakness” is workforce, which can be achieved through partnerships.
Attila Papp
Lead Cloud Solution Architect @ adesso Hungary Software
How Amazon became an AI-powered company
Amazon Inc. has set an example for retail innovation in the past two decades. Their continuously spinning “flywheel” spun out many subsequent services, which later became a key industry player (AWS, Alexa, Music, Prime, etc.). In this talk, we will examine Amazon’s use cases for Artificial Intelligence, how they push the retail industry for AI adoption, and what might be on the horizon for their ultimate AI strategy.
Itilekha Podder
European Institute of Innovation & Technology (EIT) Doctoral Student
Eötvös Loránd University (ELTE) , Budapest
Robert Bosch Kft., Budapest Hungary
AI in the MEMS based inertial sensor production
MEMS (MicroElectroMechanical System) based inertial sensors were originally used in the automotive industry in the Electronic Stability Program (ESP), but today they also can be expected to spread in the self-driving cars. The silicon technology used in MEMS sensors is developed by a series of complex and finely tuned chemical and mechanical processes. The basic elements of the products are formed with the help of silicon technology. The complexity of the process is due to the relatively large variance of the technology, which is now accompanied by the availability of well-documented process steps in the industry and the measurement results of products and semi-finished elements. The production process, the environment and the results obtained from the product allow us to deduce complex relationships, which can be reduced with the help of artificial intelligence, and their application can make certain processes more economical and more accurate. The presentation shows the possibilities of optimizing the production process with AI through examples.
Márk Jelasity
Szegedi Tudományegyetem, Természettudományi és Informatikai Kar, Informatikai Intézet, Számitógépes Algoritmusok és Mesterséges Intelligencia Tanszék
Robust Machine Learning
Deep neural networks have been known to be vulnerable to malicious input perturbations invisible to humans that can mislead the network in almost arbitrary ways. Also, the same networks are vulnerable to out-of-distribution examples, that often result in incorrect but high confidence output. In this talk, I will give a summary of some of the results we achieved in this area in the past years at the University of Szeged. This includes novel attack secnarios and algorithms, as well as formal verification.
Zoltán Kárász
Senior AI Expert at Robert Bosch
Application of AI in self-driving function development and in the product development
The development of self-driving functions today involves the parallel development of several technologies that are being addressed by several teams in parallel, including the application of Deep Learning / Machine Learning technologies. These technologies can be used in product development, but the result of such research / development is an artificial intelligence model as a product. The presentation will provide an insight into the artificial intelligence technologies, applications and their challenges used in product development and functional development of self-driving.
László Gulyás
Department of Artificial Intelligence, Faculty of Informatics, Eötvös Loránd University, Hungary
On the Structural Analysis of Sparse Neural Networks
Artificial Neural Networks (ANNs) are at the forefront of interest and practice of modern Artificial Intelligence (AI). Originally, they were proposed as a mathematical model of the human brain, i.e., the human neural network. Recent advances in ANN research resulted in a host of powerful applications, but also in novel avenues to overcome existing limitations, e.g., in terms of computational performance. One of these new directions is concerned with Sparse Neural Networks (SNNs), or with sparsification of ANNs. The core of the idea is to remove nodes or links from the generally layer-wise fully connected network topology to improve performance. A distinct, but not less important scientific discipline is Network Science (NS) that is concerned with the structural analysis of empirical networks. Since the concept of networks is a very powerful metaphor, capable of describing systems such diverse as biological, economical, social, engineered systems, among others. NS provides statistical tools to assess the structure of networks — at the micro (node), the meso (clusters) or the macro (the whole network) level. In addition, several commonalities of empirical networks were identified (e.g., they tend to be sparse), as well as structural properties that can predict system level behavior (e.g., the resilience of systems facing random failures or malicious attacks). In this presentation we introduce an approach to use the methods of NS to analyse and improve SNNs. We outline the long term potential of combining NS with SNN, but also discuss some difficulties. In addition, we present a set of baseline observations.
András Benczúr
Head of Informatics Laboratory at the Institute for Computer Science and Control, scientific coordinator of the Hungarian Artificial Intelligence National Laboratory
Embedding dynamic networks in vector spaces on-the-fly
I show methods to analyze and model network data accessible as a stream of edges for tasks including low-rank approximation, network embedding, link prediction, and centrality algorithms can be implemented and updated while the edge stream is processed. As applications, I consider topic discussions in Twitter such as COVID vaccination, as well as anonymity in cryptocurrency networks.
János Botzheim
Department of Artificial Intelligence, Faculty of Informatics, Eötvös Loránd University, Hungary, Head of Department
Spiking neural networks and their applications
As bio-inspired artificial intelligence emerges more detailed neural architectures appear. In this talk, one of the most prominent fields, the third generation of neural networks (i.e. spiking neural networks) is introduced. Along with motivations, and predicted impacts the basic modeling principles of dynamic neural networks are presented. After looking through current research problems and directions, the talk ends with example scenarios of experimental SNN applications.