Philippe Baratta: Statistical methods in observational Cosmology

Philippe Baratta: Statistical methods in observational Cosmology

Machine Learning Seminar presentation

Topic: Statistical methods in observational Cosmology

Speaker: Dr. Philippe Baratta, Aix-Marseille University and CPPM, Marseille, France

Time: 2023.01.18 at 10:00 a.m. (CET)

How to join: Please contact Jakub Lengiewicz

Format: 30′ presentation + 30′ discussion

Abstract:

Modern Cosmology is now a data science. The large amount of cosmological models aiming at describing the evolution of structures in the Universe needs to be confronted to observations. To do so, it is common in this field to resort to Bayesian statistics, although fraught with many pitfalls of interpretation. In this presentation, I will introduce the key statistical observables and tools of such an analysis. From the 3D cartography of galaxies in the universe, I will try to show that their statistical distribution is a mine of physical quantity information.

Recording:

Haralampos Hatzikirou: How can we make tumour predictions when we do not understand everything?

Haralampos Hatzikirou: How can we make tumour predictions when we do not understand everything?

Machine Learning Seminar presentation

Topic: How can we make tumour predictions when we do not understand everything?

Speaker: Haralampos Hatzikirou , Professor at the Faculty of Mathematics, Khalifa University, United Arab Emirates

Time: 2023.01.11, 10:00 a.m. CET

How to join: Please contact Jakub Lengiewicz

Format: 30′ presentation + 30′ discussion

Abstract:

In clinical reality, the need of quantitative tumour growth and progression predictions is pivotal for designing individualized therapies. To achieve this a plethora of examinations is conducted to assess the tumour lesion state, spanning from blood sample analysis, clinical imaging (e.g. CT, MRI), biopsy sampling, -omics screening etc. Such medical data correspond to snapshots in time of the patient’s state and in the current standard of care (SoC) their collection relies on patient’s clinical presentation. This implies that we cannot acquire many data timepoints hampering the personalized calibration of mathematical models and their corresponding prediction potential. Moreover, many clinical data types are not useful in informing phenotypic plasticity models hindering their clinical applicability.In a nutshell, the use of phenotypic plasticity models in the current cancer SoC faces the following challenges:  (C1) data collection is sparse in time since it relies on patient’s clinical presentation, (C2) we lack the knowledge of the precise pathways involved in regulating phenotypic plasticity mechanisms, and (C3) medical data cannot always inform mathematical models. Overcoming the afore-mentioned challenges to predict the future of a disease and propose an appropriate treatment (e.g., choice of a drug targeting proteins expressed in the tumour) is a formidable but not impossible task. In this talk I will present a novel methodology that combines mechanistic modelling and machine learning in delivering clinically relevant tumour growth predictions.

 

Additional material

Video recording: https://youtu.be/7G8irCS7RpA

Articles: https://www.nature.com/articles/s43856-021-00020-4

https://www.nature.com/articles/s41598-020-79119-y

https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1004366

Andres Posada Moreno: Concept-based explanations for convolutional neural networks

Andres Posada Moreno: Concept-based explanations for convolutional neural networks

Machine Learning Seminar presentation

Topic: Concept-based explanations for convolutional neural networks.

Speaker: Andres Posada Moreno, Institute for Data Science in Mechanical Engineering at the RWTH Aachen University

Time: Wednesday, 2022.11.30 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Format: 30′ presentation + 30′ discussion

Abstract:

Convolutional neural networks (CNNs) are increasingly being used in critical systems, where robustness and alignment are crucial. In this context, the field of explainable artificial intelligence has proposed the generation of high-level explanations of the prediction process of CNNs through concept extraction. While these methods can detect whether or not a concept is present in an image, they are unable to determine its location. What is more, a fair comparison of such approaches is difficult due to a lack of proper validation procedures. In this talk, we discuss a novel method for automatic concept extraction and localization based on representations obtained through pixel-wise aggregations of CNN activation maps. Further, we introduce a process for the validation of concept-extraction techniques based on synthetic datasets with pixel-wise annotations of their main components, reducing the need for human intervention.

Additional material

Video recording: https://youtu.be/7G8irCS7RpA

Slides:

– powerpoint slideshow: https://drive.google.com/file/d/1u_kgBZ5VEdbTjdwCfwPU6JEcrQljnQTo/view?usp=share_link

– standalone video: https://drive.google.com/file/d/1844w-9lRt-uoZCwhUN5vR7LF5vtGu0G7/view?usp=share_link

Reference paper: https://arxiv.org/abs/2206.04531

David Wagg: A time-evolving digital twin tool for engineering dynamics applications

David Wagg: A time-evolving digital twin tool for engineering dynamics applications

Machine Learning Seminar presentation

Topic: A time-evolving digital twin tool for engineering dynamics applications.

Speaker: David Wagg, Department of Mechanical Engineering, University of Sheffield

Time: Wednesday, 2022.11.23 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Abstract:

In this presentation we describe a time-evolving digital twin and its application to a proof-of-concept engineering dynamics example. The digital twin is constructed by combining physics-based and data-based models of the physical twin, using a weighting technique. The resulting model combination enables the temporal evolution of the digital twin to be optimised based on the data recorded from the physical twin. This is achieved by creating digital twin output functions that are optimally-weighted combinations of physics- and/or data-based model components that can be updated over time to reflect the behaviour of the physical twin as accurately as possible. Approximate Bayesian computation (ABC) is used for the physics-based model in this work, on the premise that relatively simple physical models are only rarely available in the context of digital twins. For the data-based model, a nonlinear auto-regressive exogeneous (NARX) neural network model was used. The engineering dynamics example is a system consisting of two cascading tanks driven by a pump. The data received by the digital twin is segmented so that the process can be carried out over relatively short time-scales. In this example, the weightings are computed based on error and robustness criteria. The results show how the time-varying water level in the tanks can be captured with the digital twin output functions, and a comparison is made with three different weighting choice criteria.

Additional material

Video recording: https://youtu.be/2nmMb_WI3zs

Orhan Ermis: A CNN-based Semi-Supervised Learning Approach for the Detection of SS7 Attacks

Orhan Ermis: A CNN-based Semi-Supervised Learning Approach for the Detection of SS7 Attacks

Machine Learning Seminar presentation

Topic: A CNN-based Semi-Supervised Learning Approach for the Detection of SS7 Attacks.

Speaker: Orhan Ermis, Luxembourg Institute of Science and Technology (LIST)

Time: Wednesday, 2022.11.09 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Over the years many standards were defined to solve the security vulnerabilities of the Signaling Systems No:7 (SS7) protocol. Yet, it still suffers from many security issues that make it vulnerable to attacks such as the disclosure of subscribers’ location, fraud, and interception of calls or SMS messages. Several security measures employ rule-based solutions for the detection of attacks in telecom core networks. However, they become ineffective when an attacker deploys sophisticated attacks that are not easily detected by filtering mechanisms. One particular solution to overcome those attacks is to use supervised machine learning solutions since they have demonstrated their ability to achieve promising results to detect anomalies in various applications. Nonetheless, they generally need to be trained with a large set of labeled data that cannot be obtained from the telecom operators due to the excessive resource allocation and cost of labeling the network traffic. Therefore, in this work, we propose an innovative approach based on semi-supervised learning, which combines the use of labeled and unlabeled data to train a particular model for the detection of SS7 attacks in telecom core networks. Our approach adapts the Convolutional Neural Network (CNN)-based semi-supervised learning scheme and an improved version of the feature engineering in previous works together with the hyperparameter optimization. Experiment results show that the proposed approach achieves up to 100\% accuracy on both the real world and simulated datasets, respectively.

Additional material

Video recording: https://youtu.be/RfzhYhbsXn8

Oscar Castro: Multi-target Compiler for the Deployment of Machine Learning Models

Oscar Castro: Multi-target Compiler for the Deployment of Machine Learning Models

Machine Learning Seminar presentation

Topic: Multi-target Compiler for the Deployment of Machine Learning Models.

Speaker: Oscar Castro, Luxembourg Institute of Science and Technology (LIST)

Time: Wednesday, 2022.10.26 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Data availability and advances in computing power nowadays have enabled a huge growth in Machine Learning (ML) research and practice. In this wave, many tools to build machine learning models have been developed. These tools are used by data scientists for fast model building and prototyping. But obtaining actual value from ML models requires their deployment into production environments. A repeatable, fast, and reliable deployment process is vital for successful ML workflows. To facilitate deployment, we have designed and developed a special-purpose compiler to automate the translation of ML models from their formal description into source code. The design of the compiler supports a dynamic architecture that allows many different types of models as inputs and many target programming languages as outputs. When compiling a ML model, we aim for running time efficiency of the deployed model on the target computer architecture. Therefore, not only we can deploy to several programming languages, but we also exploit specific underlying characteristics of the architecture such as multiple cores and the availability of graphic processing cards.

Additional material

Video recording: https://youtu.be/9acFeRw1354

Inès Chihi: Artificial intelligence is really the best solution for systems with unpredictable behavior? A comparative study between ANN and Multi-model approach

Inès Chihi: Artificial intelligence is really the best solution for systems with unpredictable behavior? A comparative study between ANN and Multi-model approach

Machine Learning Seminar presentation

Topic: Artificial intelligence is really the best solution for systems with unpredictable behavior? A comparative study between ANN and Multi-model approach.

Speaker: Inès Chihi, Department of Engineering, University of Luxembourg

Time: Wednesday, 2022.10.05 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Artificial intelligence and Data-driven models are well suggested for modeling complex and non-linear processes. However, they require a very large computation time for data preparation, analysis and also for learning. Indeed, complex problems require an extended network that can have exceptionally long and tedious computations, especially at inference instants. To overcome these problems, we propose a hybrid technique based on multi-model structure. This structure is suggested for the modeling of nonlinear process by decomposing its nonlinear operating domain into a defined number of sub-models, each one representing a well-defined operating point. Thus, the multi-model concept is considered an interesting method to improve the performance of the model in terms of accuracy and without increasing too much the complexity of the empirical model, the training time or the number of parameters to be estimated. As an example, we present a comparative study between Artificial Neural Network (ANN), known as the most used and efficient technique in empirical modeling, and the proposed multimode approach. This comparative study will be applied to estimate the muscles forces of the forearm from the muscle’s activities.

Additional material

Article: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8870134/

Video recording: https://youtu.be/TzDV88I3-uY

Samuel Renault: AI Act: a grasp of the future EU regulation on AI and its impact on AI practitioners

Samuel Renault: AI Act: a grasp of the future EU regulation on AI and its impact on AI practitioners

Machine Learning Seminar presentation

Topic: AI Act: a grasp of the future EU regulation on AI and its impact on AI practitioners.

Speaker: Samuel Renault, Luxembourg Institute of Science and Technology (LIST)

Time: Wednesday, 2022.09.28 10:00 am CET

How to join: Please contact Jakub Lengiewicz

Abstract:

The EU AI Act is a future regulation pushed by the European commission that aims at lowering and limiting the negative impact of AI on EU citizens.
After GDPR voted in 2016 and implemented in 2018, this new regulation is considered as the next game changer in the AI / ML / data science ecosystem, by making trustworthiness and accountability key principles when developing and operating AI-based systems. The regulation will largely rely on emerging AI standards to allow the regulatory text to remain valid under a dynamic evolution of the technological field.
While the regulation’s text is still under (heavy) discussion between EU institutions and interested bodies, one can get a grasp of the impact the future regulation will have on the design, development and commercialisation of AI systems in Europe and beyond.
What kind of AI systems and which organisations will be impacted by the regulation? What will be the obligations of AI system designers, operators, and users? Will it impact AI innovation in Europe, as stated by some groups of interest? These will be the topics of the talk and discussions.

Additional material

Presentation: https://legato-team.eu/wp-content/uploads/2022/09/AI-Act_Legato-ML-Webinar_20220928.pdf

Video recording: https://youtu.be/sHmkNn8BKgU

Alena Kopanicakova: Multilevel training of deep neural networks

Alena Kopanicakova: Multilevel training of deep neural networks

Machine Learning Seminar presentation

Topic: Multilevel training of deep neural networks.

Speaker: Alena Kopanicakova, Division of Applied Mathematics, Brown University, Providence, USA

Time: Wednesday, 2022.09.21 3:00 pm CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Deep neural networks (DNNs) are routinely used in a wide range of application areas and scientific fields, as they allow to efficiently predict the behavior of complex systems. However, before the DNNs can be effectively used for the prediction, their parameters have to be determined during the training process. Traditionally, the training process is associated with the minimization of a loss function, which is commonly performed using variants of the stochastic gradient (SGD) method. Although SGD and its variants have a low computational cost per iteration, their convergence properties tend to deteriorate with increasing network size. In this talk, we will propose to alleviate the training cost of DNNs by leveraging nonlinear multilevel minimization methods. We will discuss how to construct a multilevel hierarchy and transfer operators by exploring the structure of the DNN architecture, properties of the loss function, and the form of the data. The dependency on a large number of hyper-parameters will be reduced by employing a trust-region globalization strategy. In this way, the sequence of step sizes will be induced automatically by the trust-region algorithm. Altogether, this will give rise to new classes of training methods, convergence properties of which will be analyzed using a series of numerical experiments.

Additional material

Video recording: https://youtu.be/B2vPQBHR5f8

Internship offer: IsoGeometric Analysis (IGA) implementation in the open-source software SOFA

Dear all,
We are offering an internship position for 4 to 6 months at the University of Luxembourg, in the Legato team led by Prof. Stéphane P.A. Bordas (FLSW). The subject deals with implementing the IsoGeometric Analysis in the SOFA Framework. Join a dynamic team and an open-source consortium! More details are provided in the attached document.

Internship_IGA_SOFA_2022