Presentations and Publications
Selected Oral Presentations:
11:15 - Solar Flare prediction with Machine Learning based on Satellite data (Jonas Zbinden1)
[1] University of Bern
Presented by Jonas Zbinden
Abstract: With the technological advancement of our society, solar flares pose an ever-growing danger to our modern infrastructure. Yet, we still do not fully understand the physical mechanisms leading to a flare. Therefore, training machine learning models at predicting solar flares allows us to mitigate the risk for damage to our infrastructure. In our study, we investigated the potential of using spectra observed with the IRIS satellite from three different spectral lines (Mg II h&k, Si IV and C II) for solar flare forecasting. To train and test our models we used a set of 53 preflare and 30 active region observations. The model’s performances were evaluated with standard metrics, and we reach prediction accuracies up to 80%. We hypothesize that the sensitivity to heating and density changes in the mid- to high chromosphere, as well as the higher formation height range are important factors for solar flare prediction, and explain the higher performance of the Mg II h&k line compared to the other two FUV spectral lines. We also found that the biggest factor for successful flare forecasting with our models is that the IRIS slit position of the spectrograph covers the area where the flare will occur well.
11:30 - Generating Astronomical Spectra from Photometry with Conditional Diffusion Models (Lars Doorenbos1; Stefano Cavuoti2; Giuseppe Longo3; Massimo Brescia3; Raphael Sznitman1; Pablo Márquez Neila1)
[1] University of Bern
[2] Astronomical Observatory of Capodimonte
[3] University Federico II
Presented by Lars Doorenbos
Abstract: A trade-off between speed and information controls our understanding of astronomical objects. Fast-to-acquire photometric observations provide global properties, while costly and time-consuming spectroscopic measurements enable a better understanding of the physics governing their evolution. Here, we tackle this problem by generating spectra directly from photometry, through which we obtain an estimate of their intricacies from easily acquired images. This is done by using multi-modal conditional diffusion models, where the best out of the generated spectra is selected with a contrastive network. Initial experiments show promising results.
11:45 -Dynamic Federated Learning for Heterogeneous Learning Environments (Eric Samikwa1; Torsten Braun1)
[1] University of Bern
Presented by Eric Samikwa
Abstract: The emergence of the Internet of Things (IoT) has resulted in a massive influx of data generated by various edge devices. Machine learning models trained on this data can provide valuable insights and predictions, leading to better decision-making and intelligent applications. Federated Learning (FL) is a distributed learning paradigm that enables remote devices to collaboratively train models without sharing sensitive data, thus preserving user privacy and reducing communication overhead. However, despite recent breakthroughs in FL, the heterogeneous learning environments significantly limit its performance and hinder its real-world applications. The heterogeneous learning environment is mainly embodied in two aspects. Firstly, the statistically heterogeneous (usually non-independent identically distributed) data from geographically distributed clients can deteriorate the FL training accuracy. Secondly, the heterogeneous computing and communication resources in IoT devices often result in unstable training processes that slow down the training of a global model and affect energy consumption. We propose Dynamic Federated Learning (DFL) to address the joint problem of data and resource heterogeneity in FL. DFL combines resource-aware split computing of deep neural networks and dynamic clustering of training participants based on the similarity of their sub-model layers. Using resource-aware split learning, the allocation of the FL training tasks on resource-constrained participants is adjusted to match their heterogeneous computing capabilities, while resource-capable participants carry out the classic FL training. We employ centered kernel alignment for determining the similarity of neural network layers to address the data heterogeneity and carry out layerwise sub-model aggregation. Preliminary results demonstrate that DFL can significantly improve training performance (i.e., training time, accuracy, and energy consumption) in heterogeneous learning environments.
14:00 - Estimating time-dependent stratified transmission rates of SARS-CoV-2 (Judith Bouman1; Christian Althaus1; Julien Riou1; Martin Wohlfender1; Simon Grimm2; Anthony Hauser1)
[1] Institute for Social and Preventive Medicine, University of Bern
[2] Center for Space and Habitability, University of Bern
Presented by Judith Bouman
Abstract: Background: Estimating how the transmission rate of SARS-CoV-2 changes over time is critically important for accurate real-time monitoring and assessing the impact of non-pharmaceutical interventions, behavioural changes, and seasonal effects. It has been challenging to estimate time-dependent transmission rates from SIR-type models that describe infectious disease transmission across different sub-populations. Methods: We evaluated three methods (splines, Gaussian processes, and Brownian motion) to model time-dependent transmission rates for SARS-CoV-2 in a Bayesian SIR-type model implemented in Stan. Using UBELIX, we tested the performance of these methods for different implementations to optimize their efficiency. Results: Using simulated data, we show that all three methods can accurately recover time-dependent transmission rates in homogenous as well as heterogeneous populations that are stratified by age. However, we found substantial differences in their evaluation time. Applying the methods to age-stratified laboratory-confirmed SARS-CoV-2 cases from the Canton of Geneva in 2020 allowed us to obtain detailed estimates of the effective reproduction number Re over time and the ascertainment rate in each age group. Conclusions: Improved methods to estimate time-dependent transmission rates in heterogeneous SIR-type models will advance our understanding of SARS-CoV-2 transmission across different sub-populations, improve the accuracy of real-time monitoring, and inform decisions about targeted control interventions.
14:15 - Glaucoma Stage Progression Predictor (Marta Colmenar Herrera1)
[1] ARTORG Center for Biomedical Engineering Research, University of Bern
Presented by Marta Colmenar Herrera
Abstract: Glaucoma is a degenerative condition of the optic nerve that causes loss of retinal ganglion cells, leading to visual field loss and eventually to blindness. Detecting glaucoma progression early is crucial for effective management and preventing rapid sight deterioration. In this study, we develop a predictive model aimed at predicting the stage of a patient's glaucoma in their next specialist visit, using visual fields (VF), Mean Defects (MD), and the patient's age. We compared the performance of non-linear models, such as Adaboost, with linear models, like Logistic Regression, using three different datasets. Our evaluation, based on the kappa metric, revealed that Adaboost outperformed Logistic Regression with scores of 0.72 (0.02), 0.86 (0.02), and 0.74 (0.02) for the three datasets, respectively. Conversely, Logistic Regression scored 0.60 (0.02), 0.86 (0.01), and 0.64 (0.01), highlighting the linearity of the second dataset. Our results demonstrate that non-linear models can detect progression more accurately than traditionally used linear models. Therefore, our study proposes a more effective method for predicting glaucoma progression, which could help clinicians tailor treatments for patients in a timely manner.
14:30 - AI for sleep scoring: moving towards integration in the clinic (Marco Pesce1; Florence M. Aellen1; Julia Van Der Meer2; Markus Schmidt2; Cluadio Bassetti2; Athina Tzovara1)
[1] University of Bern, Bern, Switzerland
[2] Inselspital, Bern, Switzerland
Presented by Florence M. Aellen
Abstract: Not publicly available.
14:45 - Deep learning based Monte Carlo Dose Denoising for Radiation Therapy (Raphael Joost1)
[1] University of Bern
Presented by Raphael Joost
Abstract: Monte Carlo (MC) dose calculation is widely accepted as the gold-standard for calculating dose distributions (DD) of radiotherapy plans. However, its clinical applicability is limited due to its long calculation-time for MC-DD with low statistical uncertainty (SU). We developed a 3D-U-Net to denoise MC-DDs of high SU to MC-DDs of low SU for academic situations.