Norsk IKT-konferanse for forskning og utdanning

I-KAHAN: Image-Enhanced Knowledge-Aware Hierarchical Attention Network for Multi-modal Fake News Detection

2023-11-01T10:21:06+00:00

In the quest to combat the proliferation of fake news, accurate detection of fabricated news content has become increasingly desirable. While existing methodologies leverage a variety of news attributes, such as text content and social media comments, few incorporate diverse features from different modalities like images. In this paper, Image-Enhanced Knowledge-Aware Hierarchical Attention Network (I-KAHAN) architecture is proposed as an enhancement to the existing KAHAN architecture. The I-KAHAN architecture utilizes a wide variety of attributes including news content, user comments, external knowledge, and temporal information which are inherited from the KAHAN architecture, and extends it by integrating image-based information as an additional feature. This work contributes to refining and expanding fake news detection methodologies by embracing a more comprehensive range of features and modalities, and offers valuable insights into the effectiveness of various methods for the numerical representation of images, feature aggregation and dimensionality reduction. Experiments conducted on two real-world datasets, PolitiFact and GossipCop, assessing the performance of the I-KAHAN architecture, demonstrated approximately 3% improvement in accuracy over the KAHAN architecture, highlighting the potential benefits of incorporating diverse features and modalities for enhanced fake news detection performance.

Accelerating PFLOTRAN-OGS on GPUs using PETSc

2023-11-01T10:12:46+00:00

With the evident effects of rapid climate change and society´s continuing dependence on fossil fuels, efficient modelling of reservoir behaviour for emerging CO2 storage projects is in high demand. Due to the models’ physical complexity and the non-linear nature of CO2 storage processes, the computational demands are so significant that physical aspects, such as the dissolution effect, are ignored due to computational limits. State-of-the-art flow simulation codes use highly parallel hardware and programming models to handle large-domain simulations. At the time of this work, complex flow simulators had yet to fully investigate the benefits and impact of utilising accelerators like GPUs. This study explores the performance of accelerating the production code PFLOTRAN-OGS, developed by OpenGoSim, using GPUs through PETSc’s recently built-in accelerated solvers. Our accelerated simulation is run on two test cases: GW1, a CO2 storage case from OpenGoSim, and SPE1, a simple Black Oil benchmark from the Society of Petroleum Engineers. The preliminary benchmark indicates that a GPU-accelerated solver with a CPU-based framework gives an overall slower simulation. However, our profiling verifies that most of the time was spent on transferring matrices back and forth between the CPU and GPU, while the solver steps have significant speedup on the GPU. Our results thus show that the CUDA-accelerated PETSc FGMRES solver will be faster than its CPU counterpart once the complete code is moved to the GPU.

Om å kartleggja mørk materie med maskinlæring

2023-11-01T10:00:39+00:00

Gravitasjonslinsing er fenomenet der ljos frå fjerne himmellegeme vert avbøygd av tyngdekraften frå andre himmellegeme, som ofte ikkje er fullt synlege fordi mykje av massen er mørk materie. Observert gjennom ei gravitasjonslinse, framstår fjerne gallaksar som forvrengde. Der er mykje forskingsaktivitet som freistar å karleggja mørk materie ved å studera linseeffektar, men dei matematiske modellane er kompliserte og utrekningane krev i dag mykje manuelt arbeide som er svært tidkrevjande. I denne artikkelen drøftar me korleis me kan kombinera rouletteformalismen åt Chris Clarkson med maskinlæring for automatisk, lokal estimering av linsepotentialet i sterke linser, og me presenterer eit rammeverk med programvare i open kjeldekode for å generera datasett og validera resultat.

Semantics-Based Version Control for Feature Model Evolution Plans

2023-11-01T09:51:09+00:00

A software product line (SPL) models closely related software systems by capitalizing on the high similarity of the products by organizing them into common and variable parts. To ensure successful long-term development, it is beneficial to not just capture the current software product line, but the planned evolution of the SPL as well. Evolution planning of an SPL is often a dynamic, changing process due to changes in product requirements. In addition, planning is typically a collaborative effort with multiple engineers working separately and independently of each other. To improve development, their individual contributions would need to be unified. This can be a complex task, especially without proper synchronization tools. In this paper, we provide a semantics-based merge algorithm for evolution plans. Given two versions of an evolution plan and the common evolution plan they are derived from, the merge algorithm attempts to merge all the different changes from both versions. The merge algorithm will be an essential component in a version control system, allowing several contributors to unify their versions into a sound evolution plan.

Geo-locating Road Objects using Inverse Haversine Formula with NVIDIA Driveworks

2023-11-01T09:21:52+00:00

Geolocation is integral to the seamless functioning of autonomous vehicles and advanced traffic monitoring infrastructures. This paper introduces a methodology to geolocate road objects using a monocular camera, leveraging the NVIDIA DriveWorks platform. We use the Centimeter Positioning Service (CPOS) and the inverse Haversine formula to geo-locate road objects accurately. The real-time algorithm processing capability of the NVIDIA DriveWorks platform enables instantaneous object recognition and spatial localization for Advanced Driver Assistance Systems (ADAS) and autonomous driving platforms. We present a measurement pipeline suitable for autonomous driving (AD) platforms and provide detailed guidelines for calibrating cameras using NVIDIA DriveWorks. Experiments were carried out to validate the accuracy of the proposed method for geolocating targets in both controlled and dynamic settings. We show that our approach can locate targets with less than 1m error when the AD platform is stationary and less than 4m error at higher speeds (i.e. up to 60km/h) within a 15m radius.

Large-Scale Pre-Training for Dual-Accelerometer Human Activity Recognition

2023-11-01T08:59:46+00:00

The annotation of physical activity data collected with accelerometers for human activity recognition (HAR) remains challenging despite the growing interest in large public health studies. Existing free-living accelerometer-based datasets are limited, hindering the training of effective deep learning models. To address this limitation, some studies have explored self-supervised learning (SSL), i.e., training models on both labeled and unlabeled data. Here, we extend previous work by evaluating whether large-scale pre-training improves downstream HAR performance. We introduce the SelfPAB method, which includes pre-training a transformer encoder network on increasing amounts of accelerometer data (10-100K hours) using a reconstruction objective to predict missing data segments in the spectrogram representations. Experiments demonstrate improved downstream HAR performance using SelfPAB compared to purely supervised baseline methods on two publicly available datasets (HARTH and HAR70+). Furthermore, an increase in the amount of pre-training data yields higher overall downstream performance. SelfPAB achieves an F1-score of 81.3% (HARTH), and 78.5% (HAR70+) compared to the baselines' F1-scores of 74.2% (HARTH) and 63.7% (HAR70+). Additionally, SelfPAB leads to a performance increase for activities with little training data.

Forecasting Hourly Ambulance Demand for Oslo, Norway: A Neuro-Symbolic Method

2023-11-01T10:16:21+00:00

Forecasting ambulance demand is critical for emergency medical services to allocate their resources as efficiently as possible. This work uses data from Norway's Oslo University Hospital (OUH) to forecast hourly ambulance demand in Oslo and Akershus. To forecast demand, we developed a neuro-symbolic method, DeANN. DeANN integrates statistical decomposition and artificial neural network methods. Statistical decomposition computes trend, seasonal, and residual components from the ambulance demand time series. Using these components, we apply a multilayer perceptron and regression to compute an overall ambulance demand forecast. Based on experimental results, we conclude that our proposed neuro-symbolic approach for ambulance demand forecasting outperforms several baseline models. Our best neuro-symbolic model has a mean squared error of 21.68 and improves on previous results for the OUH data set.

Evolutionary Computation with Islands: Extending EvoLP.jl for Parallel Computing

2023-11-01T10:03:58+00:00

The use of evolutionary computation for optimisation is a relevant area of research in many fields of science and the industry, where complex problems are frequently encountered. As an effort to support the research in this niche, we present an extension for EvoLP.jl: the evolutionary computation playground in Julia, that includes three new operators for implementing island models for genetic algorithms. The extension enables the framework to run using the Message Passing Interface protocol, an international standard for communication in parallel architectures that is available in most high performance computing clusters today. We study the advantages of the implementation by performing a series of tests on well-known numerical optimisation benchmarks of various difficulties and on several dimensions. Both the code and the data are available in a GitHub repository. This work enables researchers to implement powerful parallel evolutionary algorithms without moving away from the high level of abstraction that the framework provides.

Simulated RGB and LiDAR Image based Training of Object Detection Models in the Context of Autonomous Driving

2023-11-01T09:55:23+00:00

The topic of object detection, which involves giving cars the ability to perceive their environment has drawn greater attention. For better performance, object detection algorithms often need huge datasets, which are frequently manually labeled. This procedure is expensive and time-consuming. Instead, a simulated environment due to which one has complete control over all parameters and allows for automated image annotation. Carla, an open-source project created exclusively for the study of autonomous driving, is one such simulator. This study examines if object detection models that can recognize actual traffic items can be trained using automatically annotated simulator data from Carla. The findings of the experiments demonstrate that optimizing a trained model using Carla’s data, along with some real data, is encouraging. The Yolov5 model, trained using pre-trained Carla weights, exhibited improvements across all performance metrics compared to one trained exclusively on 2000 Kitti images. While it didn’t reach the performance level of the 6000-image Kitti model, the enhancements were indeed substantial. The mAP0.5:0.95 score saw an approximate 10% boost, with the most significant improvement occurring in the Pedestrian class. Furthermore, it is demonstrated that a substantial performance boost can be achieved by training a base model with Carla data and fine-tuning it with a smaller portion of the Kitti dataset. Moreover, the potential utility of Carla LiDAR images in reducing the volume of real images required while maintaining respectable model performance becomes evident. Our code is available at: https://tinyurl.com/3fdjd9xb.

GECO: A Twitter Dataset of COVID-19 Misinformation and Conspiracy Theories Related to the Berlin Parliament and Washington Capitol Riots

2023-11-01T09:26:43+00:00

On August 29, 2020, a precursor to the widely known January 6 United States Capitol attack in Washington D.C., USA, occurred in Berlin, Germany, where a group of protesters participating in a demonstration against COVID-19 pandemic measures attempted to storm the German parliament in Berlin. While the event in Berlin was less dramatic than January 6 of 2021 in the US - the protesters were repelled by the police, and no serious damage or injuries were reported - in both cases, mobilization through conspiracy theories on social media is widely considered a significant factor leading to both events.

In this paper, in order to study such social media content, we present an analysis based on a manually labeled dataset sampled from a large set of COVID-19 related tweets in temporal proximity to the event in Berlin. Moreover, we provide an analysis that is based on a set of tweets following the January 6 United States Capitol event for comparison. The labels distinguish eight different classes of conspiracy theories, as well as other misinformation. This allows for studying the prevalence of different misinformation narratives around events of note. In total
23,417 tweets were labeled manually.

The purpose of this dataset analysis is to allow further study of the phenomena, as well as training of machine learning systems with the purpose of detecting conspiracy theory content.

Linear MIM-width of the Square of Trees

2023-11-01T09:17:08+00:00

Graph parameters measure the amount of structure (or lack thereof) in a graph that makes it amenable to being decomposed in a way that facilitates dynamic programming. Graph decompositions and their associated parameters are important both in practice (as a tool for designing robust algorithms for NP-hard problems) and in theory (relating large classes of problems to the graphs on which they are solvable in polynomial time).

Linear MIM-width is a variant of the graph parameter MIM-width, introduced by (Vatshelle 2012). MIM-width is a parameter that is constant for many classes of graphs. Most graph classes which have been shown to have constant MIM-width also have constant linear MIM-width. However, computing the (linear) MIM-width of graphs, or showing that it is hard, has proven to be a huge challenge. To date, the only graph class with unbounded linear MIM-width, whose linear MIM-width can be computed in polynomial time, is the trees (Høgemo et al. 2019). In this follow-up, we show that for any tree $T$ with linear MIM-width $k$, the linear MIM-width of its square $T^2$ always lies between $k$ and
$2k$, and that these bounds are tight for all $k$.

Preface

2023-11-29T14:51:21+00:00

NIK 2023