Opening

• István Szabó

Room: Mátyás Hall (Groundfloor)

Machine Learning in High-Energy Physics: Successes and Future Directions

• Graeme Stewart (CERN)

Room: Mátyás Hall (Groundfloor) Machine learning has been used in high-energy physics for several decades, with considerable success. With the advent of modern machine learning the range of applications HEP has exploded, with techniques being applied in almost every area of the field. I shall review here some of the advances and successes that have been achieved in event classification, simulation and reconstruction domains. Some indications of how success may develop in the future will be discussed as well as some practical issues related to the deployment of machine learning into the lifecycle of HEP computing.

One Year Into the ESS Control System Machine Learning Project

• Karin Rathsman (ESSS)

Room: Mátyás Hall (Groundfloor) The European Spallation Source ERIC (ESS) is a joint European organisation committed to building and operating the world's leading facility for research using neutrons. The facility design and construction includes a powerful linear proton accelerator, a helium-cooled tungsten target wheel and two dozen state-of-the-art neutron instruments. ESS is made up of a large number of diverse systems and disciplines, covering for example water cooling, vacuum, power distribution, timing systems, information technology, networking, microwaves, cryogenics among others. These systems are integrated and controlled by the central integrated control system (ICS), which acts as the “brain” of the ESS machine. The ESS machine will generate at least an order of magnitude larger volumes of control system related data than typically existing in large industries. With an estimated number 100 000 devices to control and 1.6 million process values in the control system we realize that the high-level requirement of 95% availability for the facility will be very challenging and that ICS will play a key role to reach this goal. For this reason, ESS has initiated a control system machine learning (CSML) project in 2018, with the aim to build up a collaboration between research facilities, academia and industry. The project will explore how application of modern machine learning technologies to a large-scale industrial distributed control system can help increase facility availability and efficiency and lower costs for operation. This talk will cover the activities and outcome of the first year.

Image-based classification of variable stars

• Attila Bódi (Konkoly Thege Miklós Astronomical Institute, Research Centre for Astronomy and Earth Sciences)

Room: Mátyás Hall (Groundfloor) Among the astronomical community it is well-known that different types of variable stars can be recognised based on their light curve properties. However, with the advent of large astronomical sky surveys, the amount of measurements increases dramatically, reaching the limits of the human capabilities of stellar classification. During my talk I will give a brief introduction to variable star science and give a short summary on previous efforts made to categorise light curves. After that I will present our recent results on image-based deep learning stellar classification and will discuss the possible applications in the era of big data which will produced by future surveys like the Large Synoptic Survey Telescope (LSST).

Variable star classification

• Dóra Tarczay-Nehéz (CSFK CSI)

Room: Mátyás Hall (Groundfloor) The most recent telescopes (e.g. Kepler, K2, Gaia, TESS) and sky surveys (e.g. SSDS, and the forthcoming LSST) provide huge amount of data, that leads to the challenge of data processing. This huge volume of data needs to be analyzed with fast and effective automated computer programming techniques. Therefore, machine learning algorithms become popular in astronomy, as they can play a key role in automatic classification of variable stars. In this work, we present our machine learning algorithm for searching variable stars, based on the statistical characteristics of light curves, that represent the brightness variability of the stars in the Gaia DR2 database.

Life beyond the pixels: machine learning and image analysis methods for HCS

• Péter Horváth (Hungarian Academy of Sciences)

Room: Mátyás Hall (Groundfloor) In this talk I will give an overview of the computational steps in the analysis of a single cell-based high-content screen. First, I will present a novel microscopic image correction method designed to eliminate vignetting and uneven background effects which, left uncorrected, corrupt intensity-based measurements. I will present deep learning-based image segmentation methods. I will discuss the Advanced Cell Classifier (ACC) (www.cellclassifier.org), a software tool capable of identifying cellular phenotypes based on features extracted from the image. It provides an interface for a user to efficiently train machine learning methods to predict various phenotypes. We developed the Suggest a Learner (SALT) toolbox, which selects the optimal machine learning algorithm and parameters for a particular classification problem. For cases where discrete cell-based decisions are not suitable, we propose a method to use multi-parametric regression to analyze continuous biological phenomena. Finally, to improve the learning speed and accuracy, we recently developed an active learning scheme which automatically selects the most informative cell samples.

Learning structured deep generative models from incomplete data in biomedicine

• Péter Antal (BME)

Room: Mátyás Hall (Groundfloor) Deep probabilistic generative models demonstrated superior and scalable performance in multiple domains. However, their application in biomedicine is still hindered by the following challenges, incorporation of prior knowledge, interpretation and explanation, and learning from highly incomplete data, especially from sparsely populated time-series data. At first, I illustrate standard solutions to these challenges using belief networks. Next, I overview the evolution of generative models from deep belief networks. Finally, I summarize recent extensions of deep generative models to cope with these challenges and their biomedical applications.

What can we learn about learning through human-machine interactions?

• Ferdinando Mussa-Ivaldi (Northwestern Univ and Shirley Ryan Ability Lab)

Room: Mátyás Hall (Groundfloor) A growing body of evidence suggests that when we interact physically with our environments our brains form models of the deterministic connection between our actions and the ensuing sensory information. Theories of motor learning posit that the formation of internal models is a key mechanism though which the brain forms predictions about the outcomes of actions, overcoming certain limitations of the biological feedback system. Consistent with these theories, experiments with human-robot interactions have demonstrated the ability of the brain to capture the difference between random and deterministic forces. I will review some of these earlier studies and fill focus then on a family of human-machine interfaces that create a many-to-one mapping between body motions and movements of an external controlled object. In this context, the user learns to control the external object by forming an inverse model of the interface mapping. I will describe this learning process as a state-based dynamical system and will discuss how machine learning may cooperate with human learning to facilitate the acquisition of motor skills and their recovery after injury to the nervous system.

Complete Inference of Causal Relationships in Dynamical Systems

• Zoltán Somogyvári (Wigner Research Centre for Physics, Department of Computational Sciences)

Room: Mátyás Hall (Groundfloor) Inference of causal structure between multiple observations of a complex system gained large interest in wide range of scientific disciplines, from pharmaceutics to economy, where it earned Nobel-prize for Clive Granger in 2003. In this talk, we present a new analysis method called Dimensional Causality (DC). We belive, our method is the first one, which is able to detect and distinguish all possible types of causal relations: independence, directed or circular causal connection and particularly the existence of a hidden common cause. To detect these relations between two time series, a dynamical system’s theoretical approach is combined with the Bayesian model inference. We validated our method on simulated examples of classical chaotic and non-chaotic dynamical systems such as coupled Lorentz-systems, Logistic maps, or Hindmarsh-Rose models and demonstrated its capabilities on human neurophysiological measurements. As an example of medical application, the possible focus of epileptic seizure (an area which drives the others) is identified in a patient, based on implanted electrode recordings from the surface of the brain. However, the universality of our method ensures its applicability in many other fields of science.

Hindrances to adaptation of Machine Learning in Healthcare

• Richárd Nagyfi (Cursor Insight)

Room: Mátyás Hall (Groundfloor) Artificial Intelligence is over-hyped and anthropomorphised by the media which has led to numerous unfulfilled expectations and fear-mongering throughout the decades. Despite experts prophesying omnipotent robot AGIs, there is seldom any discourse about the actual hindrances to adaptation of current narrow solutions. Black-box functionality, lack of transparency and interpretability, and misleading evaluation metrics of Machine Learning models are the main reasons why physicians remain skeptic towards the true potential of AI. Instead of jumping the hype, one should think of Machine Learning models as another set of sophisticated diagnostic tools that rely on patient data instead of the measurement of their bodily functions - another set of tools with their own use-cases and limitations. Instead of aiming to "replace doctors", the industry should instead create devices that help physicians establish more accurate diagnoses, in close collaboration.

AI ecosystem and AI strategy in Hungary

• Gergely Szertics

Room: Mátyás Hall (Groundfloor) The Minister of Innovation and Technology announced the 2020 action plan of AI in October and the elaboration of the AI strategy is in progress. We will take a look at the cornerstones of the strategy and the elements that are already announced that are focusing on both the ecosystem, the institutions and the markets that has to be developped to be able to live with the opportunities that AI brings in a controlled manner.

AI based algo-trading using quantization and volatility information

• János Levendovszky (BME)

Room: Mátyás Hall (Groundfloor) Novel algorithms are developed for algorithmic trading on financial time series by using quantization and volatility information to achieve High Frequency Trading (HFT). The proposed methods are estimation based and trading actions are carried out after estimating the Forward Conditional Probability Distribution (FCPD) on the quantized return values. For estimating FCPD, a FeedForward Neural Network (FFNN) will be used, which can provide a good estimation if the return levels are quantized properly and a special encoding scheme is applied. The performance analysis of our method tested on historical time series (NASDAQ/NYSE stocks) has demonstrated that the algorithm is profitable. As far as high frequency trading is concerned, the algorithm lends itself to GPU implementation, which can considerably increase its performance when time frames become shorter and the computational time tends to be the critical aspect of the algorithm.

Hb-graphs and their applications to the ranking of information in an information space.

• Xavier Ouvrard (CERN / University of Geneva)

Room: Mátyás Hall (Groundfloor) Hb-graphs have been previously introduced to handle redundancy in the m-uniformisation process used for the construction of an e-adjacency tensor for general hypergraphs. In this talk, we present an application of hb-graphs to co-occurrence networks of an information space, allowing a multi-diffusion scheme for ranking information.

The Hungarian National AI Center of Excellence

• Andras Benczur (SZTAKI)

Room: Mátyás Hall (Groundfloor) The initiative to found the Hungarian National AI Center of Excellence led by the Institute for Computer Science and Control (SZTAKI) was announced in October 15 this year. The Center is planned to start operation in early 2020 in the areas of basic and applied research, education, technology development as well as acting as a key Hungarian player in international cooperation. In my presentation, I will highlight our background in some planned key topics, including - Mathematical foundations of deep neural network learning; - Applications in health care, manufacturing, telecommunications, IoT, machine vision and human language technologies; - Algorithmic and legal issues of privacy.

Machine Learning and Reasoning for Exploratory Data Analysis and Model Extraction in Cyber-Physical Systems

• András Földvári

Room: Mátyás Hall (Groundfloor) Modern cyber-physical system (CPS) design relies on the paradigm of component integration. Assurance of the compliance with extra-functional requirements of critical CPS applications necessitates empirical identification before integrating components, and validation during the final acceptance test and operation, respectively. Benchmarking and operational log analysis are the primary means of validation and checking of time-related attributes, like timeliness or performability. These processes generate hard-to-interpret, many-dimensional, and big data sets. Exploratory Data Analysis (EDA) supports a better understanding of the collected data by integrating the priory knowledge of the domain expert with the observations. It extracts a model of the phenomena used later in system engineering and operational supervision of the CPS. The talk presents a novel combination of discretization, knowledge fusion, inductive logic-based automated system model identification for validating, and diagnosing complex systems. It uses a sequence of steps of quantization, knowledge base construction by merging priory knowledge and observations and inductive logic-based reasoning for generalization of the observed results for automated model extraction, consistency checking, and fault diagnosis.

Adding artificial Intelligence into process control systems – implementation and integration by open source tools

• Tamas Ruppert

Room: Mátyás Hall (Groundfloor) We demonstrate how the toolbox of artificial intelligence (AI) and machine learning (ML) can support the monitoring of processes. We highlight how these functions can be implemented in existing process control systems and how open source solutions (e.g., Python toolboxes) can be goal-oriented tailored for their development. We give an in-depth overview of the steps of the workflow of the implementation of these solutions and present the structure of an AI/ML supported process control system. The methodology and the results are presented concerning the AI/ML-based improvement of the monitoring and operation support functionalities in the WebSCADA system of an industrial water treatment plant.

Artificial Intelligence for industry at the Hartree Centre

• Andrew Gargett

Room: Mátyás Hall (Groundfloor) There would seem to be good reason to be optimistic about the uptake of Artificial Intelligence in real-world settings; despite the hype, there seems to be genuine progress with the long-standing problem of how to map areas of often obscure research to more mundane and practical challenges, at least for those organisations large enough to absorb any risks involved as well as supply the necessary resources. However, there is still concern about the level of impact for smaller commercial ventures. Our work at the Hartree Centre, within the Science and Technology Facilities Council, is a key component of a long-term goal of the UK government toward accelerating the adoption of such technology throughout UK industry. This talk will describe Hartree's experiences as we add AI, Data Science and Big Data to our established reputation as one of Europe's premier industry facing centres for High Performance Computing; particular focus will be given to our efforts to ensure we reach all levels of industry, from the largest to the smallest players, and the likely benefits to all.

Computer vision methods for understanding human movements in daily living and clinical settings

• Maura Casadio (U Genova)

Room: Mátyás Hall (Groundfloor) Measuring and understanding human motion is crucial in several domains, ranging from neuroscience, to rehabilitation and sports biomechanics. The study of human motion is commonly done through marker-based techniques and motion capture systems. If on one hand these methods are precise and reliable, on the other they present some disadvantages, in particular they are expensive, encumbering, and time consuming. For these reasons recently the research of cheaper and easier markerless techniques had made great strides. In particular, we are witnessing a steady growth in the design and implementation of computer vision and machine learning algorithms applied to plain video recordings of human movements. This type of analysis facilitates the extraction of features that give qualitative and quantitative information about human motion and that can be used for detecting, characterizing, and understanding motor behavior and deficits associated with neurological diseases. Our research team at the University of Genoa has combined multiple expertises to design and implement a markerless pipeline based on state of the art algorithms and apply it to different case studies. In this talk we will show some preliminary results.

Enhancement of 2D and 3D Medical Imaging Data acquired with Mediso SPECT Cameras

• Gábor Légrádi (physicist-engineer, AI expert, freelancer)

Room: Mátyás Hall (Groundfloor) Mediso Medical Imaging Systems is a Hungarian company which develops, manufactures and sells 3D medical imaging tools. Beside Computer Tomography (CT), Positron Emission Tomography (PET) and Magnetic Resonance Imaging (MRI) tools, Single Photon Emission Tomography (SPECT) is the leading technology of the Company. Since the very beginning of Mediso, SPECT technology has always been improved. Among many goals, development aims at three main targets: Improving 2D scintigraphy image and 3D SPECT reconstruction quality, i.e. improving clinical utility; decreasing dose burden on patients according to ALARA (As Low As Reasonably Achievable) principle; decreasing acquisition times, i.e. improving throughput and economic payback. In 2018, attention focused on the use of AI for data enhancement. In most cases, clinical SPECT tools are used for acquiring whole-body planar bone scintigraphy and bone SPECT acquisitions. In the first project, an efficient 2D denoising autoencoder neural network tool was developed and optimized for denoising whole-body planar bone scintigraphy images. For that purpose, a complex framework was built, with which a comprehensive optimization study was performed. Different architectures, training strategies and many parameters were optimized. An efficient metrics was developed for evaluating the denoiser performance both with acquisitions of normal or deteriorated statistics. Acquisitions with additional artificial lesions were also investigated. Preliminary validation has also been started with the involvement of clinical doctors who have found our solution exceptionally promising. A similar development project has also been performed aiming at the enhancement of 3D SPECT acquisition data. In this case, very special solutions must have been applied for reaching adequate convergence in terms of loss and quality of improved data. Metrics was also developed for both 3D sinograms and 3D reconstructed volumes. In our presentation, the above described developments are presented. Mediso AI group is also introduced and its tasks are briefly summarized.

HEP Graph Machine Learning for Industrial & Humanitarian Applications

• Daniel Dobos (Visiting Researcher at CERN with Lancaster University)

Room: Mátyás Hall (Groundfloor) The power of Machine Learning for graph structured data is used more and more for HEP applications like particle tracking or jet analysis. We studied different ML tracking approaches (including GNN's) used in the TrackML and HEPTrkX challenges and used a simplified game model of Nine Men's Morris to study Graph ML using quantum algorithms. While the quantum implementation is motivated by academic interest and future performance preparations, the graph network analysis and ML on it is targeted for current analysis. In combination with visual analytics it is a powerful tool to understand and analyse humanitarian partnerships, collaborations and ecosystems that we studied and will present preliminary results. Further we will discuss some selected industrial applications we are currently exploring.

Overview of applied machine learning projects at the Department of Automation and Applied Informatics

• Gabor Gulyas (BME-AUT)

Room: Mátyás Hall (Groundfloor) In this talk we cover some research areas at our department where we use machine learning as our main tool. We provide insights to these by discussing some of the key projects in details. More particularly, we discuss works in the conjunction of machine learning and data collection and analysis, image- and natural language processing. Regarding data collection and analysis, we present a pending project proposal on the privacy protection in the smart home network environment using federated learning, and we share key takeaways on how we do smart data collection in the ZalaZone tracker system. In image processing, we present our results related to enhancing medical support for heart and vascular diseases and also how we apply machine learning to control weight and quality of intermediate products in medicine manufacturing. Finally, we present a work on machine comprehension based on semantic parsing using graph-transformations.

The breakdown of photon blockade: a first order dissipative quantum phase transition (cloud-based simulation of open quantum systems)

• András Vukics (Wigner FK)

Room: Mátyás Hall (Groundfloor) First-order phase transitions characterized by the coexistence of phases are commonly observed in the surrounding world, e.g. in the freezing of water. Continuous – second-order – phase transitions also exist in classical physics, e.g. the transition between ferro- and paramagnetism at the Curie temperature. Whereas the latter class has seen straightforward generalizations to quantum systems for decades, the notion of a first-order quantum phase transition remains to be elucidated. Bistability in certain small quantum systems has been identified as signature of first order quantum phase transitions, however, this identification is problematic: a randomly switching telegraph signal between two well-resolved attractors can also be observed in quantum dynamics distinct from phase transitions. For example, the famous electron-shelving scheme – used in atomic clocks or for qubit measurement in ion-trap quantum computers – produces a similar signal without any connection to phase transitions. There is a missing element to support the interpretation of bistability as a first-order quantum phase transition: it must be shown that bistability is only a finite-size effect, and there exists an idealized thermodynamic limit, where temporal bistability is replaced by hysteresis. This idealized thermodynamic limit can be introduced such that the physical system remains a small quantum system with a few degrees of freedom, that is, the passage to the thermodynamic limit does not involve a quantum-to-classical transition. In this talk, I present a prototype of this procedure by constructing a finite-size scaling [1] for the recently-observed photon-blockade-breakdown effect [2,3] to justify its classification as a first-order dissipative quantum phase transition. The work involves heavy numerics, performed on a virtual cluster defined within the Wigner Cloud. The applied software, as always in my work, was C++QED: a versatile open-source C++/Python application-programming framework for simulating open quantum dynamics [4-7], developed by our group. The framework uses an adaptive-timestep version of the Monte Carlo wave-function method [8], which can be readily paralellized. In the talk, I will also sketch the computer-physics aspects of the work. [1] Vukics A, Dombi A, Fink J M and Domokos P 2019 Quantum 3 150 URL https://doi.org/10.22331/q-2019-06-03-150 [2] Fink J M, Dombi A, Vukics A, Wallraff A and Domokos P 2017 Phys. Rev. X 7(1) 011012 URL https://doi.org/10.1103/PhysRevX.7.011012. [3] Carmichael H J 2015 Phys. Rev. X 5(3) 031028 URL https://doi.org/10.1103/PhysRevX.5.031028. [4] http://cppqed.sf.net [5] Vukics A and Ritsch H 2007 Eur. Phys. J. D 44 585–599 [6] Vukics A 2012 Comp. Phys. Comm. 183 1381–1396 [7] Sandner R and Vukics A 2014 Comp. Phys. Comm. 185 2380–2382 [8] Kornyik M and Vukics A 2019 Comp. Phys. Comm. 238 88–101

Comparing the Markowitz Portfolio Model and the Market Graph of Pardalos

• Sándor Szabó (Univ Pécs)

Room: Mátyás Hall (Groundfloor) The Markowitz portfolio selection model is an optimization problem for investments to achieve a good return while control the risk of losses. On the technical side it is a quadratic or in other later variants a linear program. The so-called market graph is a graph based model to capture certain aspects of the inner working and structure of a stock market. More specifically it uses concepts like clique or chromatic numbers, independent sets, clique partitions from graph theory. It is clear that both models can be used in connection with a wider variety of time serieses than those of coming from the financial world. Obviously the techniques can be used in situations other than they originally were intended. In the lecture we will treat the models as general data analytic tools and carry out a detailed comparison. We will touch on the aspect of the range of applicability, computational costs, size of problems can be handled, issues of interpretation, whether they can serve as exploratory or confirmatory devices.