Hichem Omrani: Predicting air pollution using remote sensing and sensors measurement

Machine Learning Seminar presentation

Topic: Predicting air pollution using remote sensing and sensors measurement

Speaker: Hichem Omrani, Luxembourg Institute of Socio-Economic Research (LISER)

Time: Wednesday, 2021.03.03, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Air pollution is a threat to public health, having negative effects on human health and well-being. This research seminar aims to (1) predict air pollution at a fine spatial resolution using the “land use regression” model, (2) assess population exposure to air population in Luxembourg and its surrounding areas, (3) and discuss ongoing work, opportunities and challenges for future research.

Additional material:

Carlotta Montorsi: Tree based algorithms: implementing Conditional Inference Trees and Forest for old age Frailty Index predictions

Machine Learning Seminar presentation

Topic: Tree based algorithms: implementing Conditional Inference Trees and Forest for old age Frailty Index predictions.

Speaker: Carlotta Montorsi, Luxembourg Institute of Socio-Economic Research (LISER)

Time: Wednesday, 2021.02.24, 12:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Tree-based algorithms are prediction algorithms introduced by Morgan and Sonquist (1963) and popularized by Breiman et al. (1984) almost 20 years later. These algorithms aim at predicting an outcome “out of sample” based on a number of covariates. This is done by partitioning the space of the regressors in non-overlapping regions. When the task is regression, the predicted income is simply the average outcome of units reaching each terminal node. Various methods to grow trees avoiding overfitting exist: Conditional Inference Trees introduced by Hothorn et al. (2006) prevent overfitting by growing the tree conditioning the splitting to a sequence of statistical tests.

In the upcoming presentation, I will present a brief theoretical introduction to the Tree-based model with a particular focus on Conditional Inference Trees and Conditional Inference Forest. Thus, I will show its implementation on real data, namely for predicting a Frailty Index of individuals aged 50+ from different European Countries. Moreover, I will compare the predictive performance of these algorithms with other traditional ML methods and for different subsamples of the training set. Finally, I will discuss the “best predictors” identified by the most accurate of these algorithms in the different subsamples.

Additional material:

[1] Morgan, J. N., and  Sonquist,J. A. (1963). Problems in the Analysis of Survey Data, and a Proposal, Journal of the American Statistical Association, 58(302), 415– 34.
[2] Breiman, L.,Friedman,J.,Stone,C. and R. Olshen (1984). Classification and Regression Trees, Taylor & Francis, Belmont, https://doi.org/10.1201/9781315139470.
[3] Hothorn, T.,Hornik, K. and Zeileis, A. (2006). Unbiased Recursive Partitioning: A Conditional Inference Framework. Journal of Computational and Graphical Statistics.
[4] Brunori, P. and Neidhöfer, G. (2021). The Evolution of Inequality of Opportunity in Germany: A Machine Learning Approach. Review of Income and Wealth.

Minh Vu Chau: Data-driven constitutive laws for hyperelasticity modeling in eigenspace using algorithmic differentiation coupling FEniCS-Pytorch model: application to RVEs

Machine Learning Seminar presentation

Topic: Data-driven constitutive laws for hyperelasticity modeling in eigenspace using algorithmic differentiation coupling FEniCS-Pytorch model: application to RVEs

Speaker: Minh Vu Chau, FSTM, University of Luxembourg

Time: Wednesday, 2021.02.17, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Macroscale constitutive behavior of heterogeneous hyperelastic materials can be obtained from the microstructure by a homogenization process using multilevel finite elements (FE2) analysis. Although various advancements have been made, FE2 still remains a computationally expensive method. Alternatively, artificial neural networks (ANN) have been proposed as an efficient data-driven method for constitutive modeling, accepting either synthetic data from computational homogenization solutions or experimental datasets. In this presentation, we would like to introduce the Neural Network in which the constitutive relation of hyperelastic materials is achieved via computational homogenization, and is independent upon the used coordinate system.

At the microscopic scale,  a synthetic training dataset is acquired from numerical simulation of RVE using the FEniCS framework. Subsequently, a multivariable regression analysis using ANN is conducted with eigenvalues of strain-stress pairs used as inputs and outputs. After successful training, this ANN is then plugged back into the algorithmic differentiation framework by embedding Pytorch inside FeniCS which creates a coupling symbolic FEniCS-Pytorch model such that it will capture the constitutive relationship of the material. Some technical notes on how to successfully combine multiple sub-losses functions into one total loss are also discussed if we have enough time.

Additional material:

Zhiqiang Zhong: Reinforcement Learning based Meta-path Design for Heterogeneous Graph Neural Network

Machine Learning Seminar presentation

Topic: Reinforcement Learning based Meta-path Design for Heterogeneous Graph Neural Network

Speaker: Zhiqiang Zhong, FSTM, University of Luxembourg

Time: Wednesday, 2021.02.10, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Heterogeneous Information Networks (HINs), involving a diversity of node types and relation types, are pervasive in many real-world applications. Recently, increasing attention has been paid to heterogeneous graph representation learning (HGRL) which aims to embed rich structural and semantics information in HIN into low-dimensional node representations. To date, most HGRL models rely on manual customization of meta paths to capture the semantics underlying the given HIN. However, the dependency on the handcrafted meta-paths requires rich domain knowledge which is extremely difficult to obtain for complex and semantic rich HINs. Moreover, strictly defined meta-paths will limit the HGRL’s access to more comprehensive information in HINs.

To fully unleash the power of HGRL, we present a novel framework called RL-HGNN, to design different meta-paths for the nodes in a HIN. Specifically, RL-HGNN models the meta-path design process as a Markov Decision Process and uses a policy network to adaptively design a meta-path for each node to learn its effective representations. The policy network is trained with deep reinforcement learning by exploiting the performance of the model on a downstream task. We further propose an extension, RL-HGNN++, to ameliorate the meta-path design procedure and accelerate the training process. Experimental results demonstrate the effectiveness of RL-HGNN and reveal that it can identify meaningful meta-paths that would have been ignored by human knowledge.

Additional material:

Saurabh Deshpande & Milad Zeraatpisheh: Practical implementation session on Neural Networks

Machine Learning Seminar presentation

Topic: Practical implementation session on Neural Networks

Speaker: Saurabh Deshpande & Milad Zeraatpisheh, FSTM, University of Luxembourg

Time: Wednesday, 2021.02.03, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

We will create our own neural network from scratch, only using the NumPy library. Afterward, the same example will be implemented in the Keras framework. The session will also involve a Bayesian neural network approach, implemented using the TensorFlow probability library.

Additional material:

Shared code: https://colab.research.google.com/drive/1aTf490wwL7r8_eEjcW4Sw9rgPg6ntVvf#scrollTo=B1OA9W_KnmUo

Arif Sinan Uslu: The effect of self-paced neurofeedback on EEG learning: An experimental setup

Machine Learning Seminar presentation

Topic: The effect of self-paced neurofeedback on EEG learning: An experimental setup

Speaker: Arif Sinan Uslu, FHSE, University of Luxembourg

Time: Wednesday, 2021.01.27, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Neurofeedback refers to the regulation of brain activity via a brain-computer interface. In a feedback loop, the brain activity is recorded, analyzed, and fed back to the user. Iterative and dynamic feedback enables the user to learn whether or not the recorded brain activity meets a pre-specified threshold. Research in the field of cognitive electrophysiology suggests that the regulation of brain rhythms is associated with changes in cognitive functions [1,2]. However, these associations frequently apply to subgroup of participants who were able to regulate their brain activity [2,3,4]. Up until now, most studies on EEG neurofeedback applied a block design during training during which participants receive the same timed neurofeedback training blocks (e.g. 5 minutes) interspersed by rest periods. To my knowledge, there has been no study investigating the effect of self-paced neurofeedback on learning outcome although the factor of training frequency and duration has been discussed as an indicator of training success. In the current study, I will investigate how self-paced neurofeedback influences the activity of alpha frequency compared to classical externally-paced neurofeedback and a control group receiving sham-neurofeedback.

In my presentation, I will give a brief theoretical introduction to the matter and focus later on the technical implementation of my experimental setup and the type of data that is being recorded. Although a data-driven methodology is not implemented for this experiment, I would very much like to take this opportunity to discuss possibilities to include data-driven methods that could be applied to the data from this experiment or a modified version of this experiment.

Additional material:

[1] Escolano, C., Navarro-Gil, M., Garcia-Campayo, J. et al. The Effects of a Single Session of Upper Alpha Neurofeedback for Cognitive Enhancement: A Sham-Controlled Study. Appl Psychophysiol Biofeedback 39, 227–236 (2014). https://doi.org/10.1007/s10484-014-9262-9
[2] Hanslmayr, S., Sauseng, P., Doppelmayr, M., Schabus, M., & Klimesch, W. (2005). Increasing individual upper alpha power by neurofeedback improves cognitive performance in human subjects. Applied psychophysiology and biofeedback, 30(1), 1–10. https://doi.org/10.1007/s10484-005-2169-8
[3] Nan, W., Rodrigues, J. P., Ma, J., Qu, X., Wan, F., Mak, P. I., Mak, P. U., Vai, M. I., & Rosa, A. (2012). Individual alpha neurofeedback training effect on short term memory. International journal of psychophysiology : official journal of the International Organization of Psychophysiology, 86(1), 83–87. https://doi.org/10.1016/j.ijpsycho.2012.07.182
[4] Kober, S. E., Witte, M., Ninaus, M., Neuper, C., & Wood, G. (2013). Learning to modulate one’s own brain activity: the effect of spontaneous mental strategies. Frontiers in human neuroscience, 7, 695. https://doi.org/10.3389/fnhum.2013.00695 (edited) 

Milad Zeraatpisheh: Bayesian neural networks and MC Dropout; ways to measure uncertainty in Deep learning

Machine Learning Seminar presentation

Topic: Bayesian neural networks and MC Dropout; ways to measure uncertainty in Deep learning

Speaker: Milad Zeraatpisheh, FSTM, University of Luxembourg

Time: Wednesday 2021.01.20, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Deep learning methods represent the state-of-the-art for numerous applications such as facial recognition systems, supercomputing, and speech recognition. Conventional Neural networks generate point estimates of deep neural network parameters and therefore, make predictions that can be overconfident since they do not account well for uncertainty in model parameters.

In this presentation, we will take a closer look at the Bayesian Neural network as a way to measure this uncertainty. First, Bayesian inference on Neural network weights will be discussed. Afterward, Monte Carlo Dropout, proposed by Gal & Ghahramani (2016), as another way to tackle uncertainty in deep learning will be explained.

Additional material:

 

Lars Beex: Unsupervised learning to select modes for reduced-order models of hyperelastoplasticity: application to RVEs

Machine Learning Seminar presentation

Topic: A Physics-guided Machine Learning Model Based on Peridynamics

Speaker: Dr Lars Beex, FSTM, University of Luxembourg

Authors: S. Vijayaraghavan (1; 2), L.A.A.Beex (1), L.Noels (2), and S.P.A.Bordas (1)

1) University of Luxembourg, Ave. de la Fonte 6, L-4364 Belval, Luxembourg.

2) University of Liege, CM3 B52, Alle de la découverte 9, B4000 Liege, Belgium

Time: Wednesday 2021.01.13, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

Many model order reduction approaches use solutions of a few ‘offline’ training simulations to reduce the number of degrees of freedom of the many ‘online’ simulations of interest. In proper orthogonal decomposition, singular value decomposition is applied to a matrix with the training solutions in order to capture the most essential characteristics in the first few modes – which are used as global interpolation bases.

Proper orthogonal decomposition has proven itself as an accurate reduced-order model approach for elliptical partial differential equations. In the field of solid mechanics, this means that it is accurate for (hyper)elastic material models, but not for (hyper)elastoplasticity. Based on the study of [1], the current contribution investigates how clustering of the training solutions and extracting global modes from each cluster can improve the accuracy of proper orthogonal decomposition for hyperelastoplasticity. Both centroid-based clustering (i.e. k-means clustering) and connectivity-based clustering (based on Chinese whispers) are investigated.

The approach is applied to a hyperelastoplastic representative volume element exposed to monotonic loading, quasi-monotonic loading, and quasi-random loading. In case of monotonic and quasi-monotonic loading, the components of the macroscopic deformation tensor are the variables to which clustering is applied. In case of quasi-random loading, however, not only the components of the macroscopic deformation tensor and the incremental changes of these components are the variables to which clustering is applied, but also all history variables of all integration points.

Additional material:

[1] David Amsallem, Matthew J. Zahr and Charbel Farhat. Nonlinear model order reduction based on local reduced-order bases. Int. J. Numer. Meth. Engng. VOL 92 IS-10 SN-0029-5981

Erkan Oterkus: A Physics-guided Machine Learning Model Based on Peridynamics

Machine Learning Seminar presentation

Topic: A Physics-guided Machine Learning Model Based on Peridynamics

Speaker: Pofessor Erkan Oterkus, PeriDynamics Research Center, University of Strathclyde, Glasgow, UK

Time: Wednesday 2020.12.16, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

With the rapid growth of available data and computing resources, using data-driven models is a potential approach in many scientific disciplines and engineering. However, for complex physical phenomena that have limited data, the data-driven models are lacking robustness and fail to provide good predictions. Theory-guided data science is the recent technology that can take advantage of both physics-driven and data-driven models. In this webinar, a new physics-guided machine learning model based on peridynamics will be presented. Peridynamics is a suitable approach for predicting progressive damages because the theory uses integro-differential equations instead of partial differential equations. Several numerical examples will be shown to demonstrate the capability of the methodology.

Additional material:

 

Eleni Koronaki: From partial data to out-of-sample parameter and observation estimation with Diffusion Maps and Geometric Harmonics

Machine Learning Seminar presentation

Topic: From partial data to out-of-sample parameter and observation estimation with Diffusion Maps and Geometric Harmonics

Speaker: Dr. Eleni Koronaki, University of Luxembourg, Department of Computational Science

Time: Wednesday 2020.12.09, 10:00 CET

How to join: Please contact Jakub Lengiewicz

Abstract:

A data-driven framework is presented, that enables the estimation of quantities, either observations or parameters, given enough partial data. Here the data are high-dimensional vectors containing the outputs of a detailed CFD model of the process, i.e. the values of velocity, pressure, temperature and species mass fractions at each point in the discretization. The goal is to predict the outputs of new inputs, here process parameters and also the inputs that correspond to new outputs.

Finally, along the lines of a nonlinear observer, part of the outputs are predicted given only a limited number of partial observations, i.e. not the entire output vector but rather a handful of independent values. The proposed workflow begins by determining the intrinsic reduced description of the available data with Diffusion Maps, a data-driven nonlinear manifold learning technique that can be thought of as the nonlinear counterpart of Principal Components Analysis. This low-dimensional description of the high dimensional ambient space that contains the data, is then leveraged for efficient interpolation and regression with a special implementation of Geometric Harmonics.

Additional material: