ABSTRACT. That software moves the world is a clear fact. And that it is becoming more and more important, too. There are three aspects that have led to an increase in the intensity with which software is used: the Internet and social networks, data and artificial intelligence. However, not everything is positive in the support that software provides to our daily lives. There are estimates that ICT will be responsible for 20% of global energy consumption by 2030, part of which will be due to software. And precisely the three mentioned aspects require large amounts of energy. In this keynote we will review different concepts related to software sustainability, and we will show some results of software consumption measurements that we have carried out on the one hand, cases carried out to raise awareness in society in general about the impact that software has on the environment. On the other hand, examples related to the consumption of data and artificial intelligence carried out with the aim of creating a set of best practices for the software professionals. Our ultimate goal is to make you aware of the consumption problem associated with software and to ensure that, if at first, we were concerned with the “what” and later with the “how”, now it is time to focus on the “with what”.

ABSTRACT. Physics-Informed Neural Networks (PINNs) are a promising application of deep neural networks for the numerical solution of nonlinear partial differential equations (PDEs). However, it has been observed that standard PINNs may not be able to accurately fit all types of PDEs, leading to poor predictions for specific regions in the domain. A common solution is to partition the domain by time and train each time interval separately. However, this approach leads to the prediction errors being accumulated over time, which is especially the case when solving “stiff” PDEs. To address these issues, we propose a new PINN training scheme, called DP-PINN (Dual-Phase PINN). DP-PINN divides the training into two phases based on a carefully chosen time point ts. The phase-1 training aims to generate the accurate solution at ts, which will serve as the additional intermediate condition for the phase-2 training. New sampling strategies are also proposed to enhance the training process. These design considerations improve the prediction accuracy significantly. We have conducted the experiments to evaluate DP-PINN with both “stiff” and non-stiff PDEs. The results show that the solutions predicted by DP-PINN exhibit significantly higher accuracy compared to those obtained by the state-of-the-art PINNs in literature.

ABSTRACT. The issue of data-driven neural network model construction is one of the core problems in the domain of Artificial Intelligence. A standard approach assumes a fixed architecture with trainable weights. A conceptually more advanced assumption is that we not only train the weights, but also find out the optimal model architecture. In this paper, we present a new method that realizes just that. We show how to create a neural network with a procedure that allows dynamic shrinking and growing of the model while it is being trained. The decision-making mechanism for the architectural design is governed by a Monte Carlo tree search procedure which simulates network behavior and allows to comparison of several candidate architecture changes to choose the best one. The solution utilizes a Stochastic Gradient Descent-based optimizer developed from scratch to realize the task of network architecture modification. The paper is accompanied with a Python source code of the prepared method. The proposed approach was tested in visual pattern classification problems and yielded highly satisfying results.

ABSTRACT. The Chilean coast is a very seismically active region. In the 21st century, the Chilean region experienced 19 earthquakes with a magnitude of 6.2 to 8.8, where 597 people were killed. The most dangerous earthquakes occur at the bottom of the ocean. The tsunamis they cause are very dangerous for residents of the surrounding coasts. In 2010, as many as 525 people died in a destructive tsunami caused by an underwater earthquake. The purpose of our research paper is to develop a tsunami simulator based on the modern methodology of Physics Informed Neural Networks (PINN) . We test our model using a tsunami caused by a hypothetical earthquake off the coast of the densely populated area of Valparaiso, Chile. We employ a longest-edge refinement algorithm expressed by graph transformation rules to generate a sequence of triangular computational meshes approximating the seabed and seashore of the Valparaiso area based on the Global Multi-Resolution Topography Data available. For the training of the PINN, we employ points from the vertices of the generated triangular mesh.

ABSTRACT. Physics-informed neural networks (PINNs) have been successfully applied to several problems in science and engineering: fluid dynamics, medical imaging, material science, geophysics, climate modelling, wave propagation, and inverse problems, among others. Despite its success, they sometimes fail to converge to the solution. Several remedies have been applied to this problem, but one of the most promising is the dynamical adaptation of the collocation points. This paper explores a novel adaptive sampling method based on the Adaptive Mesh Refinement used in the Finite Element Method. We test our method in a 1D advection-dominated diffusion problem that exhibits a steep solution near the right boundary as the diffusion coefficient tends to zero with promising results.

ABSTRACT. We propose a single-rate method for geometry compression of triangle meshes based on using a neural predictor to predict the encoded vertex positions using connectivity and an already known part of the geometry. The method is based on standard traversal-based methods but uses a neural predictor for prediction instead of a hand-crafted prediction scheme. The parameters of the neural predictor are learned on a dataset of existing triangle meshes. The method additionally includes an estimate of the prediction uncertainty, which is used to guide the encoding traversal of the mesh. The results of the proposed method are compared with a benchmark method on the ABC dataset using both mechanistic and perceptual metrics.

ABSTRACT. Interior point methods are widely used for different types of mathematical optimization problems. Many implementations of interior point methods in use today rely on direct linear solvers to solve systems of equations in each iteration. The need to solve ever larger optimization problems more efficiently and the rise of hardware accelerators for general purpose computing has led to a large interest in using iterative linear solvers instead, with the major issue being inevitable ill-conditioning of the linear systems arising as the optimization progresses. We investigate the use of Krylov solvers for interior point methods in solving optimization problems from radiation therapy and support vector machines. We implement a prototype interior point method using a so called doubly augmented formulation of the Karush-Kuhn-Tucker linear system of equations, originally proposed by Forsgren and Gill, and evaluate its performance on real optimization problems from radiation therapy and support vector machines. Crucially, our implementation uses a preconditioned conjugate gradient method with Jacobi preconditioning internally. Our measurements of the conditioning of the linear systems indicate that the Jacobi preconditioner improves the conditioning of the systems to a degree that they can be solved iteratively, but there is room for further improvement in that regard. Furthermore, profiling of our prototype code shows that it is suitable for GPU acceleration, which may further improve its performance in practice. Overall, our results indicate that our method can find solutions of acceptable accuracy in reasonable time, even with a simple Jacobi preconditioner.

ABSTRACT. We define the EC (Elliptic Curve)-based factorization witnesses and prove related results within both conditional and unconditional approaches. We present experimental computations that support the conjecture of behavior of related admissible elliptic curves in relation to the deterministic complexity of suitable factoring algorithms based on the parameters of the witnesses. This paper features three main results devoted to the factorization of RSA numbers $N = pq$, where $q>p$. The first result of computational complexity of elliptic curve factorization is improvment by the factor $D^{\sigma}$, comparing to previously known result $O\left( D^{(2+o(1))}\right)$, where $D$ is smoothness bound, assuming additional knowledge of the admissible elliptic curve. The second result demonstrates the feasibility of achieving factorization in deterministic, polynomial time, based on knowledge obtained at a specific step in the elliptic curve method (ECM), a feat previously considered impossible. The third result establishes that deterministic time for conditional factorization using the elliptic Fermat method is equal to a polynomial function of $\left(\frac{q}{p}\left(1+\left(\frac{|a_p|+|a_q|}{D}\right)^2\right)\right)^{1+o(1)}$, where $a_p,a_q$ are the Frobenius traces of the corresponding curves ($E(\mathbb{F}{p}), E(\mathbb{F}_{q})$), and $D$ indicates the approximation of the quotient $p/q$ by the quotient of $a_p/a_q$, assuming the order of the group of points over a pseudo elliptic curve $E(\mathbb{Z}_{N})$ is known.

ABSTRACT. The tensor-matrix multiplication is a basic tensor operation required by various tensor methods such as the ALS and the HOSVD. This paper presents flexible high-performance algorithms that compute the tensor-matrix product according to the Loops-over-GEMM (LoG) approach. Our algorithms can process dense tensors with any linear tensor layout, arbitrary tensor order and dimensions all of which can be runtime variable. We discuss different tensor slicing methods with parallelization strategies and propose six algorithm versions that call BLAS with subtensors or tensor slices. Their performance is quantified on a set of tensors with various shapes and tensor orders. Our best performing version attains a median performance of 1.37 double precision Tflops on an Intel Xeon Gold 6248R processor using Intel's MKL. We show that the tensor layout does not affect the performance significantly. Our fastest implementation is on average at least 14.05% and up to 3.79x faster than other state-of-the-art approaches and actively developed libraries like Libtorch and Eigen.

ABSTRACT. One of the consequences of climate change is the increase in forest fires around the world. In order to act quickly when this type of natural disaster occurs, it is important to have simulation tools that allow a better approximation of the evolution of the fire, especially in Wildland Urban Interface (WUI) areas. Most forest fire propagation simulators tend to represent the perimeter of the fire in a polygonal way, which often does not allow us to capture the real evolution of the fire in complex environments, both at the terrain and vegetation levels. In this work, we focus on Elliptical Wave Propagation (EWP) based simulators, which represent the perimeter of the fire with a set of points connected to each other by straight lines. When the perimeter grows and new points must be added, the interpolation method used is linear interpolation. This system generates unrealistic shapes of fires. In this work, an interpolation method leveraging Composite Bézier Curves (CBC) is proposed to generate fire evolution shapes in a more realistic way. The proposed method has been incorporated into FARSITE, a well-known EWP-based forest fire spread simulator. Both interpolation methods have been applied to ideal scenarios and a real case. The results show that the proposed interpolation method (CBC) is capable of generating more realistic fire shapes and, in addition, enables the simulator the ability to better simulate the spread of fire in WUI zones.

ABSTRACT. This paper introduces a computational framework for optimizing vegetation removal, modelled via so-called blackline or fireline widths, to enhance efficiency and cost-effectiveness of a prescribed burn for planned reduction of vegetation density. The QUIC FIRE simulation tool is employed to conduct simulations across fireline widths ranging from 8 to 24 meters in 2-meter increments within a strategically chosen burn unit that covers the usecase of a wildland urban interface located around the region of Auburn, CA. Through visual analysis and quantitative cost function assessment, incorporating polynomial fit and the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm within a basin-hopping framework, an optimal fireline width is computed that minimizes costs, efforts and the risk of fire escapes. Findings indicate that strategic adjustments in fireline widths significantly influence the success of prescribed burns, underscoring the value of advanced simulation and optimization techniques. This work provides a foundational framework for subsequent studies, advocating for the development of dynamic, adaptive models that are scalable across varied ecological and geographical settings. Contributions extend to a computational and economic perspective on sustainable risk mitigation, underlining the pivotal influence of technology and advanced modeling in the evolution of prescribed burn strategies.

ABSTRACT. This paper addresses the challenge of classifying objects based on fragmented data, particularly when dealing with characteristics extracted from images captured from various angles. The complexity increases when dealing with fragmentary images that may partially overlap. The paper introduces a classification model utilizing neural networks, specifically multilayer perceptron (MLP) networks. The key concept involves generating local models based on local tables comprising characteristics extracted from fragmented images. Since the local tables may have different sets of attributes due to varying perspectives, missing attributes in the tables are imputed by introducing artificial objects. The local models, now with identical structures are created and the aggregation of these models into a global model is carried out using weighted averages. The model's efficacy is evaluated against existing literature methods using various metrics, demonstrating superior performance in terms of F-measure and balanced accuracy. Notably, the paper investigates the impact of the number of generated artificial objects on classification quality, revealing that a higher number generally improves results.

ABSTRACT. The determination of three-dimensional positions and velocities of stars based on the observations collected by a space telescope suffers from the uncertainty of random as well as systematic errors. The systematic errors are introduced by imperfections of the telescope’s optics and detectors as well as in the pointing accuracy of the satellite. The fine art of astrometry consists of heuristically finding the best possible calibration model that will account for and remove these systematic errors. Since this is a process based on trial and error, appropriate software is needed that is efficient enough to solve the system of astrometric equations and reveal the astrometric parameters of stars for the given calibration model within a reasonable time. In this work, we propose a novel architecture and corresponding prototype of a direct solver optimized for running on supercomputers. The main advantages expected of this direct method over an iterative one are the numerical robustness, accuracy of the method, and the explicit calculation of the variance-covariance matrix for the estimation of the accuracy and correlation of the unknown parameters. This solver is supposed to handle astrometric systems with billions of equations within several hours. To reach the desired performance, state-of-the-art libraries for parallel computing are used along with the hand-crafted subroutines optimized for hybrid parallelism model and advanced vector extensions of modern CPUs. The developed solver is tested using the mock science data related to the Japan Astrometry Satellite Mission for INfrared Exploration (JASMINE).

ABSTRACT. In recent years, self-supervised learning has played a pivotal role in advancing machine learning by allowing models to acquire meaningful representations from unlabeled data. An intriguing research avenue involves developing self-supervised models within an information-theoretic framework, but many studies often deviate from the stochasticity assumptions. Our research demonstrates that by adhering to these assumptions, specifically by employing stochastic embeddings in the form of a parametrized conditional density, we can not only achieve performance comparable to deterministic networks but also significantly improve the detection of out-of-distribution examples, surpassing even the performance of supervised detectors. Remarkably, this improvement is achieved solely by leveraging information from the underlying embedding distribution.

ABSTRACT. Anomaly detection is a critical aspect of uncovering unusual patterns in data analysis. This involves distinguishing between normal patterns and abnormal ones, which inherently involves uncertainty. This paper presents an enhanced version of the framework parallel UC2B for anomaly detection, previously introduced in a different context. In this work, we present an extension of the framework and present its large-scale evaluation on the Supercomputer Fugaku. The focus is on assessing its scalability by leveraging a great number of nodes to process large-scale datasets within the cybersecurity domain, using the UNSW-NB15 dataset. The ensemble learning techniques and inherent parallelizability of the Unite and Conquer approach are highlighted as key components, contributing to the framework's computational efficiency, scalability, and accuracy. This study expands upon the framework's capabilities and emphasizes its potential integration into an existing Security Orchestration, Automation, and Response (SOAR) system for enhancing cyber threat detection and response.

ABSTRACT. Organisational routines are repetitive, recognisable patterns of interdependent action by human and digital actors to accomplish tasks. Routine Dynamics is a theoretical base that informs discussion and analysis of such routines. We note that there is a knowledge gap in the literature on organisational routines to consolidate the ontologies of routines into an abstract data model. In this paper we design and implement a data model of routines using the Unified Modelling Language (UML). We present a demonstration of real-world instances routines to illustrate our data model’s use, and how one can then use instantiations of the model to analyse and simulate organisational routines based on real-world data. This example examines patterns of action inferred from data in the GitHub issue tracking system about the open-source software (OSS) project, Scikit-learn. Our study extends the theoretical/empirical understanding and knowledge base of Routine Dynamics by laying the groundwork towards examining organisational routines from a modeldriven perspective that gives rise to simulating routine dynamics.

ABSTRACT. This paper presents a student project carried out in collaboration with a major industry partner, demonstrating the simultaneous novel application of predictive analytics, in particular machine learning (ML), in the domain of boxed meal services, and explores its implications for IT education. Drawing from a validated ML model trained on data collected from box meal companies, this study showcases how predictive analytics can accurately predict customer sociodemographic characteristics, thereby facilitating targeted marketing strategies and personalized service offerings. By elucidating the methodology and results of the ML model, this article demonstrates the practical utility of computational techniques in real-world electronic services. Moreover, it discusses the pedagogical implications of incorporating such case studies into computational science education, highlighting the opportunities for experiential learning, interdisciplinary collaboration, and industry relevance. Through this exploration, the article contributes to the discourse on innovative teaching methodologies in computational science, emphasizing the importance of bridging theory with practical applications to prepare students for diverse career pathways in the digital era.

ABSTRACT. HPC education has become essential in recent years, especially when parallel computing on high performance computing systems enables modern machine learning models to grow at scale. The significant increase in computational power of the most recent supercomputers relies on a large number of cores, among others, of modern CPUs and GPUs. As a consequence, parallel program development based on parallel thinking has become a necessity for full utilization of modern HPC systems' computational power. Therefore, teaching high performance computing has become essential in developing skills demanded by industry. In this work we share our experience from conducting a dedicated HPC course, we provide a brief description of the course content and we propose a way to conduct HPC laboratory classes, in which a single task is implemented using several APIs, i.e., MPI, OpenMP, CUDA as well as hybrid MPI+Pthreads and MPI+OpenMP. Based on the actual task of verification of Goldbach's conjecture for a given range of numbers, we present and analyze the performance evaluation of students' solutions and code speed-ups for MPI and OpenMP. Additionally, we have evaluated students' subjective assessment of ease-of-use of particular APIs along with lengths of codes and students' performance over recent years.

ABSTRACT. Teacher evaluation in the classroom is a multifaceted challenge, with no one-size-fits-all solution. While student evaluations provide valuable feedback, they are limited by students' inability to accurately assess their own learning. This reliance on student evaluations alone can lead to biased assessments, grade inflation, and misconceptions about learning. To address these issues, a multidimensional ap-proach to teacher evaluation is proposed. This approach incorporates various as-sessment methods, including on-site evaluations, standardized tests, portfolio re-views, and student surveys, to provide a comprehensive view of a teacher's per-formance

ABSTRACT. Grade inflation is a complex phenomenon observed predominantly in institutions like the Tecnológico de Monterrey and English-speaking universities, with spo-radic instances in certain European regions. Its causes are multifaceted, ranging from competition among universities for prestigious opportunities to shifts in teaching and evaluation methods. Efforts to curb grade inflation, such as limiting grade distribution and publishing course averages, have had mixed success. We carry out both ChatGPT led and traditional bibliographical research to understand its causes, identify successful strategies, and adopt a collaborative, systemic ap-proach that can effectively control grade inflation without compromising teaching standards or exacerbating student stress

ABSTRACT. In one calculation, adjoint sensitivity analysis provides the gradient of a quantity of interest with respect to all system's parameters. Conventionally, adjoint solvers need to be implemented by differentiating computational models, which can be a cumbersome task and is code-specific. To propose an adjoint solver that is not code-specific, we develop a data-driven strategy. We demonstrate its application on the computation of gradients of long-time averages of chaotic flows. First, we deploy a parameter-aware echo state network (ESN) to accurately forecast and simulate the dynamics of a dynamical system for a range of system's parameters. Second, we derive the adjoint of the parameter-aware ESN. Finally, we combine the parameter-aware ESN with its adjoint version to compute the sensitivities to the system parameters. We showcase the method on a prototypical chaotic system. Because adjoint sensitivities in chaotic regimes diverge for long integration times, we analyse the application of ensemble adjoint method to the ESN. We find that the adjoint sensitivities obtained from the ESN match closely with the original system. This work opens possibilities for sensitivity analysis without code-specific adjoint solvers.

ABSTRACT. Typically, for analysing and modelling social phenomena, networks are a convenient framework that allows for the representation of the interconnectivity of individuals. These networks are often considered transmission structures for processes that happen in society, e.g. diffusion of information, epidemics, and spread of influence. However, constructing a network can be challenging, as one needs to choose its type and parameters accurately. As a result, the outcomes of analysing dynamic processes often heavily depend on whether this step was done correctly. In this work, we advocate that it might be more beneficial to step down from the tedious process of building a network and base it on the level of the interactions instead. By taking this perspective, we can be closer to reality, and from the cognitive perspective, human beings are directly exposed to events, not networks. But we can also draw a parallel to stream data mining, bringing a valuable apparatus for stream processing. Apart from taking the interaction stream perspective as a typical way in which we should study social phenomena, this work advocates that it is possible to map the concepts embodied in human nature and cognitive processes to the ones that occur in interaction streams. Exploiting this mapping can help reduce the diversity of problems that one can find in data stream processing for machine learning problems. Finally, we demonstrate one of the use cases in which the interaction stream perspective can be applied, namely, the social learning process.

ABSTRACT. Cloud-Edge Continuum is an innovative approach that exploits the strengths of the two paradigms: Cloud and Edge Computing. This new approach gives us a holistic vision of this environment enabling new kinds of applications that can exploit both the Edge Computing advantages (e.g. real-time response, data security and so on) and the powerful Cloud Computing infrastructure for high computational requirements. In this paper, we propose a Cloud-Edge Computing Workflow solution for Machine Learning (ML) inference in a hydrogeological use case. Our solution is designed in a Cloud-Edge Continuum environment thanks to Pegasus Workflow Management System Tools that are exploited for the implementation phase. The proposed work splits the inference tasks distributing the computation performed by each layer between Cloud and Edge infrastructure in a transparent way. We use two models to implement a proof-of-concept of the proposed solution.

ABSTRACT. Understanding how emissions from point sources affect the atmospheric concentrations of Greenhouse Gases (GHG) locally and on a wider scale is crucial to quantify their impact on climate change. To this end, different ways of performing global monitoring of GHG concentration using remote sensing data have been explored. The main difficulty remains to find the right balance between high resolution monitoring, which is often incomplete, and global monitoring, but at a coarser resolution. This study proposes the application of Super Resolution, a Deep Learning technique commonly employed in Computer Vision, to increase the resolution of atmospheric CO2 L3 satellite data. The resulting maps are achieving an approximate resolution of 1km*1km and are then compared with a benchmark of existing methods, before being used for emissions monitoring.

ABSTRACT. This presentation delves into the intersection of Machine Learning, Dynamical Systems, and Algorithmic Information Theory (AIT), exploring the connections between these areas. In the first part, we focus on Machine Learning and the problem of learning kernels from data using Sparse Kernel Flows. We draw parallels between Minimum Description Length (MDL) and Regularization in Machine Learning (RML), showcasing that the method of Sparse Kernel Flows offers a natural approach to kernel learning. By considering code lengths and complexities rooted in AIT, we demonstrate that data-adaptive kernel learning can be achieved through the MDL principle, bypassing the need for cross-validation as a statistical method.   Transitioning to the second part of the presentation, we shift our attention to the task of simplifying Partial Differential Equations (PDEs) using kernel methods. Here, we utilize kernel methods to learn the Cole-Hopf transformation, transforming the Burgers equation into the heat equation. We argue that PDE simplification can also be seen as an MDL and a compression problem, aiming to make complex PDEs more tractable for analysis and solution. While these two segments may initially seem distinct, they collectively exemplify the multifaceted nature of research at the intersection of Machine Learning, Dynamical Systems, and AIT, offering preliminary insights into the synergies that arise when these fields converge.

ABSTRACT. In this paper a multi-domain multi-task algorithm for feature selection in bulk RNAseq data is proposed. Two datasets are investigated arising from mouse host immune response to Salmonella infection. Data is collected from several strains of collaborative cross mice. Samples from the spleen and liver serve as the two domains. Several machine learning experiments are conducted and the small subset of discriminative across domains features have been extracted in each case. The algorithm proves viable and underlines the benefits of across domain feature selection by extracting new subset of discriminative features which couldn’t be ex- tracted only by one-domain approach.

ABSTRACT. We propose a new graphical model to describe the comorbidity of allergic diseases. We present our model in two versions. First, we introduce a generative model that reflects the variables’ causal relationship. Then, we propose an approximation of the generative model by a misspecified model, which is computationally more efficient and easily interpretable. We focus on the misspecified version, which we consider more practical. In both versions of our model, we consider typical allergic disease symptoms and covariates. We consider two directed acyclic graphs (DAGs). The first one describes information about the coexistence of certain allergic diseases (binary variables). The second graph describes the relationship between particular symptoms and the occurrence of these diseases. In the generative model, the edges lead from diseases to symptoms, corresponding to causal relations. In the misspecified model, we reverse the direction of edges: they lead from symptoms to diseases. The proposed model is evaluated on a cross-sectional multicentre study in Poland on the ECAP database (www.ecap.pl). An assessment of the stability of the proposed model is obtained using bootstrap and jackknife techniques.

ABSTRACT. Image recognition techniques heavily rely on abundant labeled data, particularly in medical contexts. Addressing the challenges associated with obtaining labeled data has led to the prominence of self-supervised learning and semi-supervised learning, especially in scenarios with limited annotated data. In this paper, we proposed an innovative approach by integrating self-supervised learning into semi-supervised models to enhance medical image recognition. Our methodology commences with pre-training on unlabeled data utilizing the BYOL method. Subsequently, we merge pseudo-labeled and labeled datasets to construct a neural network classifier, refining it through iterative fine-tuning. Experimental results on three different datasets demonstrate that our approach optimally leverages unlabeled data, outperforming existing methods in terms of accuracy for medical image recognition.

ABSTRACT. Transmembrane proteins (TMPs) are crucial to cell biology, making up about 30% of all proteins based on genomic data. Despite their importance, most of the available software for aligning protein sequences focuses on soluble proteins, leaving a gap in tools specifically designed for TMPs. Only a few methods target TMP alignment, with just a couple of the available to researchers. Considering that there are a few particular differences that ought to be taken into consideration aligning TMPs sequences, standard MSA methods are ineffective to align TMPs. In this paper, we present TM-MSAligner, a software tool designed to deal with the multiple sequence alignment of TMPs by using a multi-objective evolutionary algorithm. Our software include features such as transmembrane substitution matrix dynamically used according to the topology region, a high penalty to gap opening and extending, and two MSA quality scores, Sum-Of-Pairs with Topology Prediction and Aligned Segments, that can be optimized at the same time. This approach reduce the number of Transmembrane (TM) and non-Transmembrane (non-TM) broken regions and improve the TMP quality score. TM-MSAligner outputs the results in an HTML format, providing an interactive way for users to visualize and analyze the alignment. This feature allows for the easy identification of each topological region within the alignment, facilitating a quicker and more effective analysis process for researchers.

ABSTRACT. Started at NASA as a ‘living model’ to mitigate the Apollo 13 oxygen tank explosion, the concept of a digital twin where a computational model of a physical system takes continues input data from that system, predicting future events based on that data, and if needed intervening with the physical system, has witnessed a strong growth over the last decade. I will report on a vision on Digital Twins in Healthcare (DTH) emerging in Europe, and discuss the example of a DTH for Acute Ischemic Stroke. In doing so I will also reflect on recent trends in Computational Science as triggered o.a. by developments in Digital Twins.

ABSTRACT. Three-dimensional point clouds play a fundamental role in a wide array of fields, spanning from computer vision to robotics and autonomous navigation. Modeling the 3D shape of objects from these point clouds is important for various applications, including 3D shape completion and object recognition. This paper presents a complex-valued product-unit network for data-driven 3D shape completion. Using product units, sparse superpositions of complex power laws, including sparse polynomial functions, are fitted to incomplete 3D point clouds and used for extrapolating the data in the 3D space. In computer-vision applications, this task occurs frequently, e.g., when only partial views are available or occlusions hinder the acquistion of the full point cloud. We conduct a comparative analysis with a standard neural network to emphasize the superior extrapolation capabilities of product-unit networks within the 3D space. Furthermore, we present a real-world task that serves as a tangible demonstration of the proposed method's utility in the context of completing incomplete point cloud data acquired with a 3D scanner. This research contributes new insights into the field of neural network applications for 3D point cloud processing, revealing the broad potential of product-unit networks in this domain.

ABSTRACT. Traditional 3D face models are based on mesh representations with texture. One of the most important models is Flame (Faces Learned with an Articulated Model and Expressions), which produces meshes of human faces that are fully controllable. Unfortunately, such models have problems with capturing geometric and appearance details. In contrast to mesh representation, the neural radiance field (NeRF) produces extremely sharp renders. However, implicit methods are hard to animate and do not generalize well to unseen expressions. It is not trivial to effectively control NeRF models to obtain face manipulation.

The present paper proposes a novel approach, named NeRFlame, which combines the strengths of both NeRF and Flame methods. Our method enables high-quality rendering capabilities of NeRF while also offering complete control over the visual appearance, similar to Flame.

In contrast to traditional NeRF-based structures that use neural networks for RGB color and volume density modeling, our approach utilizes the Flame mesh as a distinct density volume. Consequently, color values exist only in the vicinity of the Flame mesh. This Flame framework is seamlessly incorporated into the NeRF architecture for predicting RGB colors, enabling our model to explicitly represent volume density and implicitly capture RGB colors.

ABSTRACT. Recently proposed methods for reconstructing 3D objects use a mesh with an atlas consisting of planar patches that approximate the object's surface. However, in real-world scenarios, the surfaces of reconstructed objects exhibit discontinuities that degrade the mesh's quality. Therefore, conducting additional research on methods to overcome discontinuities and improve mesh quality is always advantageous. This paper proposes to address the limitation by maintaining local consistency around patch vertices. We present LoCondA, a Locally Conditioned Atlas that represents a 3D object hierarchically as a generative model. The model initially maps the point cloud of an object onto a sphere and subsequently enforces the mapping to be locally consistent on both the sphere and the target object through the use of a spherical prior. Using this method, the mesh can be sampled on the sphere and then projected back onto the manifold of the object, yielding diverse topologies that can be seamlessly connected. The experimental results demonstrate that this approach produces structurally coherent reconstructions with meshes of comparable quality to those of competitors.

ABSTRACT. This work investigates the feasibility of applying Neural Ra- diance Fields (NeRFs) for reconstructing 3D representations of damaged structures caused by the ongoing aggression of Russia against Ukraine. The drone footage depicting the devastation was utilized and three NeRF models, Instant-NGP, Nerfacto, and SplatFacto, were employed. The models were evaluated across various damage levels (0: no damage, 4: high damage) using visual quality metrics (SSIM, LPIPS, PSNR) and rendering speed metrics (FPS, NRS). No clear correlation was observed between damage level and reconstruction quality metrics, suggesting these metrics might not be reliable indicators of damage severity. SplatFacto consistently achieved the highest rendering speed (FPS, NRS) and exhibited the best visual quality (SSIM, PSNR, LPIPS) across all damage levels. Nerfacto offered a balance between speed and quality, while Instant-NGP exhibited the slowest rendering speeds. The findings suggest that NeRFs, particularly SplatFacto, hold promise for rapid reconstruction and visualization of damaged structures, potentially aiding in damage assessment, documentation, and cultural heritage preservation efforts. Moreover, the study sheds light on the po- tential applications of such advanced modeling techniques in archiving and documenting conflict zones, providing a valuable resource for fu- ture investigations, humanitarian efforts, and historical documentation. However, further research is needed to explore the generalizability and robustness of NeRFs in diverse real-world scenarios.

ABSTRACT. This paper investigates the impact of various modifications introduced to current state-of-the-art Convolutional Neural Network (CNN) architectures specifically designed for the steganalysis of digital images. Usage of deep learning methods has consistently demonstrated improved results in this field over the past few years, primarily due to the development of newer architectures with higher classification accuracy compared to their predecessors. Despite the advances made, further improvements are desired to achieve even better performance in this field. The conducted experiments provide insights into how each modification affects the classification accuracy of the architectures, which is a measure of their ability to distinguish between stego and cover images. Based on the obtained results, potential enhancements are identified that future CNN designs could adopt to achieve higher accuracy while minimizing their complexity compared to current architectures. The impact of modifications on each model’s performance has been found to vary depending on the tested architecture and the steganography embedding method used.

ABSTRACT. The classical logit dynamic on a continuous action space for decision-making under uncertainty is generalized to the dynamic where the exponen-tial function for the softmax part has been replaced by a rational one that in-cludes the former as a special case. We call the new dynamic as the rational logit dynamic. The use of the rational logit function implies that the uncer-tainties have a longer tail than that assumed in the classical one. We show that the rational logit dynamic admits a unique measure-valued solution and the solution can be approximated using a finite difference discretization. We also show that the vanishing-noise limit of the rational logit dynamic exists and is different from the best-response one, demonstrating that influences of the uncertainty tail persist in the rational logit dynamic. We finally apply the rational logit dynamic to a unique fishing competition data that has been re-cently acquired by the authors.

ABSTRACT. Clustering in presence of uncertainty may be considered, at the same time, to be a pressing need and a challenge to effectively address many real-world problems. This concise literature review aims to identify and discuss the associated body of knowledge according to a cross-domain perspective. A semi-systematic methodology has allowed the selection of 68 papers, with a priority on the most recent contributions. The analysis has re-marked the relevance of the topic and has made explicit a trend to domain-specific solutions over generic-purpose approaches. On one side, this trend enables a more specific set of solutions within specific communities; on the other side, the resulting distributed approach is not always well-integrated in the mainstream and may generate a further fragmentation of the body of knowledge, mostly because of some lack of abstraction in the definition of specific problems. While these gaps are largely understandable within the research community, a lack of implementations to provide ready-to-use resources is overall critical, looking at a more and more computational and data intensive world.

ABSTRACT. The numerical solution in sense of Prager&Synge is defined as a hypersphere containing a true solution of a system of partial differentiation equations (PDE). In the original variant Prager&Synge method is based on special orthogonal properties of PDE and may be applied only to several equations. Herein, the Prager&Synge solution (center and radius of the hypersphere) is estimated using the ensemble of numerical solutions obtained by independent algorithms for resolving of an arbitrary system of partial differentiation equations. This approach is not problem dependent and may be applied to arbitrary system of PDE. Several options for computation of the Prager&Synge solution are considered. The first one is based on the search for the orthogonal truncation errors and their transformation. The second is based on the orthogonalization of approximation errors obtained using the defect correction me-thod. It applies the superposition of numerical solutions. The third option uses the width of the ensemble of numerical solutions. The numerical tests for the two dimensional inviscid flows are presented that demonstrate the acceptable effectivity of the approximation error estimates based on the solution in the Prag-er&Synge sense.

ABSTRACT. Advances in natural language processing techniques, such as named entity recognition and normalization to widely used standardized terminologies like UMLS or SNOMED-CT, along with the digitalization of electronic health records, have significantly advanced clinical text analysis. This study presents ClinLinker, a novel approach employing a two-phase pipeline for medical entity linking that leverages the potential of in-domain adapted language models for biomedical text mining: initial candidate retrieval using a SapBERT-based bi-encoder and subsequent re-ranking with a cross-encoder, trained by following a contrastive-learning strategy to be tailored to medical concepts in Spanish. This methodology, focused initially on content in Spanish, substantially outperforming multilingual language models designed for the same purpose. This is true even for complex scenarios involving heterogeneous medical terminologies and being trained on a subset of the original data. Our results, evaluated using top-k accuracy at 25 and other top-k metrics, demonstrate our approach's performance on two distinct clinical entity linking Gold Standard corpora, DisTEMIST (diseases) and MedProcNER (clinical procedures), outperforming previous benchmarks by 40 points in DisTEMIST and 43 points in MedProcNER, both normalized to SNOMED-CT codes. These findings highlight our approach's ability to address language-specific nuances and set a new benchmark in entity linking, offering a potent tool for enhancing the utility of digital medical records. The resulting system is of practical value, both for large scale automatic generation of structured data derived from clinical records, as well as for exhaustive extraction and harmonization of predefined clinical variables of interest.

ABSTRACT. This study investigates the application of Large Language Models (LLMs) in generating commentaries on neuroscientific papers, with a focus on their stylometric differences from human-written texts. Utilizing three seminal papers from the field of medical neuroscience, each accompanied by published expert commentaries, we compare these with commentaries generated by state-of-the-art LLMs. Through quantitative stylometric analysis and qualitative assessments, we aim to identify distinguishing features and assess the viability of LLMs in augmenting scientific discourse within the domain of medical neuroscience.

ABSTRACT. Abstract. Background – The emergence of generative AI tools, em- powered by Large Language Models (LLMs), has shown powerful capa- bilities in generating content. To date, the assessment of the usefulness of such content, generated by what is known as prompt engineering, has become an interesting research question. Objectives – Using the mean of prompt engineering, we assess the similarity and closeness of such contents to real literature produced by scientists. Methods – In this exploratory analysis, (1) we prompt-engineer ChatGPT and Google Bard to generate clinical content to be compared with literature coun- terparts, (2) we assess the similarities of the contents generated by com- paring them with counterparts from biomedical literature. Our approach is to use text-mining approaches to compare documents and associated bigrams and to use network analysis to assess the terms’ centrality. Re- sults – The experiments demonstrated that ChatGPT outperformed Google Bard in cosine document similarity (38% to 34%), Jaccard docu- ment similarity (23% to 19%), TF-IDF bigram similarity (47% to 41%), and term network centrality (degree and closeness). We also found new links that emerged in ChatGPT bigram networks that did not exist in lit- erature bigram networks. Conclusions – The obtained similarity results show that ChatGPT outperformed Google Bard in document similarity, bigrams, and degree and closeness centrality. We also observed that Chat- GPT offers linkage to terms that are connected in the literature. Such connections could inspire asking interesting questions and generate new hypotheses.

ABSTRACT. Recently, the pharmaceutical industry has been integrating cutting-edge technology, such as generative AI and quantum computing, to boost drug development research efficiency. This work advances this line of research by developing a hybrid quantum-classical generative AI model to design ligands for the 1SYH protein, linked to neurological and mental problems. The model leverages a D-Wave quantum annealer to assist the training process, motivated by recent proposals that suggest annealers as a statistical source for post-classical distributions for use in generative AI methods. After training, the quality of the ligand candidates generated by the quantum-classical model is assessed using standard drug validity criteria and reward scores obtained via computational biochemistry methods. The results show that the model provides a large number of potentially efficient ligands for the 1SYH protein, displaying chemical diversity and introducing novel molecules that meet the demanding parameters for computational success rate. Future efforts will focus on scaling up the model to larger sizes, allowing for the embedding of a broader range of chemical compounds.

ABSTRACT. Variational Data Assimilation (DA) is a technique aimed at mitigating the error in simulated states by integrating observations. Variational DA is widely employed in weather forecasting and hydrological modeling as an optimization technique for refining dynamic simulation states. However, when constructing the cost function in variational DA, it is necessary to establish a transformation function from simulated states to observations. When observations come from ground sensors or from remote sensing, representing such a transformation function with explicit expressions can sometimes be challenging or even impossible. Therefore, considering the strong mapping capabilities of Neural Network (NN)s in representing the relationship from simulated states to observations, this paper proposes a method utilizing a NN as the transformation function. We evaluated our method on a real dataset of river discharge in the UK and achieved a 13% enhancement in prediction accuracy, measured by Mean Square Error (MSE), compared to the results obtained without DA.

ABSTRACT. The detection and explanation of anomalies within the industrial context remains a difficult task, which requires the use of welldesigned methods. In this study, we focus on evaluating the performance of Explainable Anomaly Detection (XAD) algorithms in the context of a complex industrial process, specifically cold rolling. We train several state-of-the-art anomaly detection algorithms on the synthetic data from the cold rolling process and optimize their hyperparameters to maximize its predictive capabilities. Then we employ various model-agnostic Explainable AI (XAI) methods to generate explanations for the abnormal observations. The explanations are evaluated using a set of XAI metrics specifically selected for the anomaly detection task in industrial setting. The results provide insights into the impact of the selection of both machine learning and XAI methods on the overall performance of the model, emphasizing the importance of interpretability in industrial applications. For the detection of anomalies in cold rolling, we found that autoencoderbased approaches outperformed other methods, with the SHAP method providing the best explanations according to the evaluation metrics used.

ABSTRACT. This study addresses persistent challenges in breast cancer diagnosis by introducing a novel Convolutional Neural Network (CNN) approach that seamlessly integrates medical image segmentation and classification. The segmentation process demonstrates remarkable precision, delineating Normal, Benign, and Malignant regions with Jaccard Index (JI) values of 0.89, 0.92, and 0.87, contributing to an overall JI of 0.896. High Dice Similarity Coefficient (DSC) values, notably for Benign (0.94), Normal (0.96), and Malignant (0.92) regions, underscore the model's accuracy in segmenting both non-cancerous and cancerous areas, yielding an overall DSC of 0.943. Across all classes, the CNN model exhibits high accuracy, specificity, precision, recall, and F1 score, establishing its reliability for clinical applications. This research not only confronts the challenges in breast cancer diagnostics but also proposes an innovative CNN-based method as a solution. The comprehensive numerical evaluation of the model's performance metrics enhances understanding of its potential clinical utility. Beyond its immediate applications, this study lays a robust foundation for future accurate and personalized medical imaging advancements. It provides an effective tool for clinicians to enhance diagnostic accuracy and improve patient outcomes.

ABSTRACT. Computational modeling emerged to address scientific problems by developing mathematical models for their description and creating computational codes to obtain solutions. Its use in studying cardiac electrophysiology has led to a better understanding of heart function, requiring considerable time and technological expertise. MonoWeb is a structured platform for simulating electrophysiological activity in cardiac tissues, using the monodomain model in an entirely browser-based way. This tool provides a comprehensive and accessible platform for modeling and analyzing cardiac electrical activity, integrating visualization and flexible configuration in an intuitive interface. Through communication with the MonoAlg3D simulator, it allows the input of advanced parameters, and different cellular models, including selecting arrhythmia examples, stimuli, and even tissue sizes, all through the browser with just a few clicks, with the main goal being turning this experience easier and practical for cardiologists and other professionals interested in electrophysiology.

ABSTRACT. Background: Since the launch of various generative AI tools, scientists have been striving to evaluate their capabilities and contents, in the hope of establishing trust in their generative abilities. Regulations and guidelines are emerging to verify generated contents and identify novel uses.

Objective: We aspire to demonstrate how ChatGPT claims are checked computationally using the rigor of network models. We aim to achieve fact-checking of the knowledge embedded in biological graphs that were contrived from ChatGPT contents at the aggregate level.

Methods: We adopted a biological networks approach that enables the systematic interrogation of ChatGPT's linked entities. We designed an ontology-driven fact-checking algorithm that compares biological graphs constructed from approximately 200,000 PubMed abstracts with counterparts constructed from a dataset generated using the ChatGPT-3.5 Turbo model.

Results: In 10-samples of 250 randomly selected records a ChatGPT dataset of 1000 "simulated" articles, the fact-checking link accuracy ranged from 70% to 86%. The computational process was followed by a manual process using IntAct Interaction database and the Gene regulatory network database (GRNdb) to confirm the validity of the links identified computationally. We also found that the proximity distance (90 to 153) of the edges of ChatGPT graphs was significantly shorter than the literature distance (236 to 765). This pattern held true in all 10-samples.

Conclusion: This study demonstrated high accuracy of aggregate disease-gene links relationships found in ChatGPT-generated texts. The strikingly consistent pattern illuminates of new biological pathways that may open the door for new research opportunities.

ABSTRACT. The study of the relationship between brain neuron impulses and the behavioral responses of humans and other animals is an area of interest, but the one that has received relatively little attention in scientific biology and medical research centers. Advances in neurorecording technologies have enabled the simultaneous measurement of the activity of thousands of brain cells, but the challenge of analyzing the resulting data remains significant. This paper considers the problem of determining a mouse position in a ring based on neural activity data of the mouse brain. This task, which is one of the more challenging problems in this field, can be addressed through several approaches. This paper investigates the use of machine learning to solve this problem. The study is conducted in two variations: as a classification task and as a regression task. The paper presents the formulation of each problem and describes the methods used to analyze the data and make predictions. The results of the study are discussed, and the implications for future research are presented. For each of the scenarios, a solution is proposed by constructing a convolutional neural network based on the graph of brain neurons activity. In the first scenario, the model predicts the sector of the ring that contains the mouse position coordinates at the particular moment of time. In the second scenario, it predicts the exact mouse position coordinates for each time step. For each task, accuracy results were achieved: 89% for the classification task and 93% for the regression task.