Telenor (Norway)

The computing continuum aims to break the isolation of edge and cloud computing, creating a smooth and heterogeneous infrastructure surface for deploying applications. However, the continuum is practically fragmented, with infrastructure managed in isolation by various parties in a vertical dimension, i.e., edge, fog, and cloud layers, and horizontally, i.e., computing, network, and storage domains. This scenario negatively impacts the fulfillment of the application objectives. We propose inCoord, an intent-aware solution that enables the creation of a unified computing continuum by coordinating its instances to fulfill application objectives. Each instance represents a cluster of (compute, storage, or network) nodes with their own manager component. Traditionally, systems take low-level actions on these instances. In contrast, inCoord learns their emerging behaviors and adapts their managers’ objectives to fulfill the application’s intents. Here, through a Reinforcement-Learning-based Proof of Concept, we show the potential of this system to understand emerging behavior and manage multi-domain, independently managed instances.

Oxidative stress destabilizes microRNA homeostasis in the retinal pigment epithelium (RPE), driving apoptosis and the epithelial-to-mesenchymal transition, which contribute to age-related macular degeneration. We investigated whether Quantum Molecular Resonance (QMR®) electrostimulation, alone or combined with Patient Blood-Derived (PBD) secretoma, can reprogram the RPE miRNome and mitigate stress-induced damage. Human ARPE-19 cells were exposed to tert-butyl-hydroperoxide and treated with QMR®, PBD secretome, or their combination. The deep sequencing of small RNAs at 24 h and 72 h, followed by differential expression and pathway enrichment analyses, delineated treatment-driven miRNA signatures. Oxidative stress deregulated > 50 miRNAs, enriching pro-apoptotic, fibrotic, and inflammatory pathways. QMR® restored roughly 40% of these miRNAs and upregulated additional cytoprotective species such as miR-590-3p, a known regulator of the NF-κB and NLRP3 pathways according to validated target databases. While these observations suggest the potential involvement of inflammatory and stress-related cascades, functional assays will be required to directly confirm such effects. Secretome treatment preferentially increased anti-inflammatory miR-146a-5p and regenerative miR-204-5p while suppressing pro-fibrotic let-7f-5p. Combined QMR® + secretome triggered the broadest miRNA response, normalizing over two-thirds of stress-altered miRNAs. These changes are predicted to influence antioxidant, anti-apoptotic, and anti-fibrotic pathways, although they did not translate into additional short-term cytoprotection compared with QMR® alone. These data indicate that QMR® and PBD secretome modulate complementary miRNA programs that converge on stress response networks. This broader molecular reprogramming may reflect regulatory complementarity, but functional validation is needed to determine whether it provides benefits beyond those observed with QMR® alone. These findings offer molecular insights into potential non-invasive, cell-free strategies for retinal degeneration, although in vivo validation will be required before any clinical translation to Age-Related Macular Degeneration (AMD) therapy.

In Federated Learning, heterogeneity in client data distributions often means that a single global model does not have the best performance for individual clients. Consider for example training a next-word prediction model for keyboards: user-specific language patterns due to demographics (dialect, age, etc.), language proficiency, and writing style result in a highly non-IID dataset across clients. Other examples are medical images taken with different machines, or driving data from different vehicle types. To address this, we propose a simple yet effective personalized federated learning framework (pFedLIA) that utilizes a computationally efficient influence approximation, called ‘Lazy Influence’, to cluster clients in a distributed manner before model aggregation. Within each cluster, data owners collaborate to jointly train a model that captures the specific data patterns of the clients. Our method has been shown to successfully recover the global model’s performance drop due to the non-IID-ness in various synthetic and real-world settings, specifically a next-word prediction task on the Nordic languages as well as several benchmark tasks. It matches the performance of a hypothetical Oracle clustering, and significantly improves on existing baselines, e.g., an improvement of 17% on CIFAR100.

In this paper we investigate people’s attitudes to privacy and sharing of personal data when used to help society combat a contagious disease, such as COVID-19. Through a two-wave survey, we investigate the role of personal characteristics, and the effect of information, in shaping privacy attitudes. By conducting the survey in Norway and Sweden, which adopted very different strategies to handle the COVID-19 pandemic, we analyze potential differences in privacy attitudes due to policy changes. We find that privacy concern is negatively correlated with allowing public use of personal data. Trust in the entity collecting data and collectivist preferences are positively correlated with this type of data usage. Providing more information about the public benefit of sharing personal data makes respondents more positive to the use of their data, while providing additional information about the costs associated with data sharing does not change attitudes. The analysis suggests that stating a clear purpose and benefit for the data collection makes respondents more positive about sharing. Despite very different policy approaches, we do not find any major differences in privacy attitudes between Norway and Sweden. Findings are also similar between the two survey waves, suggesting a minor role for contextual changes.

This paper proposes and investigates a reconfigurable intelligent surface (RIS) enabled bidirectional non-orthogonal multiple access (NOMA) network termed as NOMA-RIS. Here, RIS allows multiple NOMA users in one group to communicate with or share information with multiple NOMA users in another group. Specifically, the two NOMA user groups send the data intended for exchange to the RIS. RIS reflects the NOMA signals allowing bidirectional communication between two NOMA user groups. In particular, under the Rician fading environment, we pay close attention to how well RIS-enabled bidirectional NOMA networks operate. Analytical expressions for tight upper bounds for the ergodic capacity are mathematically derived and verified with the simulation results. Comprehensive performance comparisons are presented, showing that our proposed bidirectional NOMA-RIS system can achieve enhanced capacity gains than RIS-enabled traditional or-thogonal multiple access schemes. These comparisons also offer practical insights into the impact of various system parameters on the overall network performance.

We present a novel time series anomaly detection method that achieves excellent detection accuracy while offering a superior level of explainability. Our proposed method, TimeVQVAE-AD, leverages masked generative modeling adapted from the cutting-edge time series generation method known as TimeVQVAE. The prior model is trained on the discrete latent space of a time-frequency domain. Notably, the dimensional semantics of the time-frequency domain are preserved in the latent space, enabling us to compute anomaly scores across different frequency bands, which provides a better insight into the detected anomalies. Additionally, the generative nature of the prior model allows for sampling likely normal states for detected anomalies, enhancing the explainability of the detected anomalies through counterfactuals. Our experimental evaluation on the UCR Time Series Anomaly archive demonstrates that TimeVQVAE-AD significantly surpasses the existing methods in terms of detection accuracy and explainability. We provide our implementation on GitHub: https://github.com/ML4ITS/TimeVQVAE-AnomalyDetection.

The visions of 5G and Beyond (B5G) imply unprecedented expectations toward high-performing connectivity services in both public and private networks. Connectivity services that offer performance guarantees along multiple Quality of Service (QoS) dimensions are partially available today, but are confined to (virtual) private network services. However, open and equal access to public and Internet-scale Specialized Connectivity Services (SCS) delivered on-demand does not exist. This especially true for interconnections across networks and support for mixed traffic modes that go beyond traditional best-effort. In this paper, we argue that this is a huge industrial and societal problem that needs a solution. However, this problem is highly complex and multi-faceted, and there are many reasons why we are essentially locked into the status quo. We identify the stumbling blocks and propose a set of solution elements to take us across these hurdles, alongside related research topics. This includes an approach to "Multi-Level Best-Effort (MLBE)" and suggestions for evolving net neutrality regulation. Models and simulations show how a mixed traffic mode approach provides anticipated benefits, and we provide arguments why the context brought by B5G will put us into conditions for change, allowing public SCS eventually at a global scale.

The diffusion of the Fifth-Generation (5G) of<br> mobile radio networks will be the main driver in the digital<br> transformation towards a new hyper-connected society. In<br> order to satisfy the stringent demands of 5G-ready applications<br> over the limited resources available at the edge, scaling<br> mechanisms become crucial to guarantee the performance levels<br> envisaged for 5G. Such mechanisms must be able to<br> automatically perform according to the real-time user demands,<br> the availability of computing resources and the state of Network<br> Functions (NFs) and applications. In this context, this paper<br> proposes a deep learning model, based on Artificial Neural<br> Networks (ANNs), for the dynamic and automated<br> orchestration of NFs. The novelty of this model is its<br> independence from specific 5G NF implementations; this is due<br> to the nature of the Key Performance Indicators (KPIs) used in<br> this work, which are related to both execution environment<br> (standard “black-box” KPIs) and standard 5G APIs (“whitebox”<br> KPIs). Results obtained on the orchestration of a Session<br> Management Function (SMF) reach an accuracy of 97~98% for<br> the training and validation phases and above 95% for the<br> deployed model, as well as higher overall accuracy by ~5% and<br> computational resource savings with respect to a thresholdbased<br> scheme.

Using the GAMBIT global fitting framework, we constrain the MSSM with an eV-scale gravitino as the lightest supersymmetric particle, and the six electroweakinos (neutralinos and charginos) as the only other light new states. We combine 15 ATLAS and 12 CMS searches at 13 TeV, along with a large collection of ATLAS and CMS measurements of Standard Model signatures. This model, which we refer to as the {{tilde{G}}}-EWMSSM, exhibits quite varied collider phenomenology due to its many permitted electroweakino production processes and decay modes. Characteristic {{tilde{G}}}-EWMSSM signal events have two or more Standard Model bosons and missing energy due to the escaping gravitinos. While much of the {{tilde{G}}}-EWMSSM parameter space is excluded, we find several viable parameter regions that predict phenomenologically rich scenarios with multiple neutralinos and charginos within the kinematic reach of the LHC during Run 3, or the High Luminosity LHC. In particular, we identify scenarios with Higgsino-dominated electroweakinos as light as 140 GeV that are consistent with our combined set of collider searches and measurements. The full set of {{tilde{G}}}-EWMSSM parameter samples and GAMBIT input files generated for this work is available via Zenodo.