| Keynote: Ilir Fetaii, SBB |
| Title: ML Model != ML Application |
| Abstract: The goal of this presentation is to provide an understanding of the essential components and success factors of Machine Learning (ML) applications through a concrete practical example. As an ML Engineer, I have learned to develop models and optimize model parameters, but that is just the tip of the iceberg. ML applications encompass much more than just model development. At SBB, we have had the privilege of guiding several ML projects to the point of being production-ready. The insights presented in this talk are based on the experiences we gained during the implementation of these projects. |
| Authors: Nima Akbari and Isabel Wagner |
| Title: Privacy and Transparency for Virtual Reality |
| Abstract: Virtual Reality (VR) headsets offer immersive games and productivity applications. However, technical features of modern VR headsets, such as eye- and face-tracking via inward-facing cameras, give rise to privacy concerns. For example, users’ emotional reactions to VR content could be leaked to third parties including advertisers. In this paper, we highlight existing works that study privacy and transparency for virtual reality, show what open gaps and challenges remain, and explain how we will address these in future work. |
| Authors: Rahel Arnold and Heiko Schuldt |
| Title: Exploring Mixed Reality Query Generation for Enhanced Multimedia Retrieval |
| Abstract: This report presents a novel concept that merges mixed reality (MR) technology with database queries for multimedia retrieval and offers a research prototype application for iOS. Our new MR approach detects real-world objects used in queries to retrieve similar multimedia from a connected database. Through seamless integration of object recognition, query formulation, and database search, users interact intuitively with their environment. This approach offers dynamic and immersive user experiences, presenting a promising research direction for future information retrieval. |
| Authors: Yiming Wu, Heiko Schuldt and Marco Vogt |
| Title: Integrating DuckDB into Polypheny-DB |
| Abstract: In this project, we envision the integration of DuckDB into Polypheny. This integration has the potential to expand the Polypheny ecosystem and create new possibilities for integrating Polypheny with data analysis pipelines. Our approach involves utilizing DuckDB’s JDBC driver and AbstractJdbcStore. Nonetheless, challenges may arise during the implementation of this integration, potentially rendering certain operations unfeasible. Addressing these challenges will be the primary focus of our upcoming research phase. |
| Authors: Tim De Ryck, Ameya D. Jagtap and Siddhartha Mishra |
| Title: On the stability of extended and conservative physics-informed neural networks (XPINNs and cPINNs) |
| Abstract: Physics-informed neural networks (PINNs) are an attractive framework in scientific machine learning that allows to approximate solutions to PDEs without any training data and in high dimensions. We theoretically investigate the stability of two PINN variants, extended PINNs (XPINNs) and conservative PINNs (cPINNs), by verifying whether a small XPINN/cPINN loss corresponds to a good approximation of the PDE. We provide counterexamples that show that this does not need to be the case for XPINNs and prove that minimizing the cPINN loss function indeed leads to good PDE approximations. |
| Authors: Jonas Linkerhägner, Cheng Shi and Ivan Dokmanić |
| Title: Interpreting Graph Denoising via Resonance |
| Abstract: Graph Neural Networks (GNNs) are an effective method to extract information from the structure and the features of a graph. To achieve strong performance, most works either 1) design a specific GNN architecture to control the way information propagates on the graph or 2) rewire the graph itself to enhance the extraction of information. In this paper we show that there is a strong analogy between graph denoising and the physical phenomenon of resonance. Similar to identifying resonant frequencies of dynamical systems, both the design of the GNN and rewiring denoise the graph by matching the graph signal with an eigenvector of the graph’s Laplacian. Preliminary numerical experiments show that aiming for resonance by jointly denoising the signal and the graph can improve the performance of GNNs for node classification tasks. |
| Authors: Monika Nagy-Huber and Volker Roth |
| Title: Physics-Informed Boundary Integral Networks (PIBI-Nets): Data Assimilation for the Heat Equation |
| Abstract: The heat equation is one of the most popular partial differential equations (PDEs) with numerous applications from physics and engineering to environmental sciences. Data assimilation models became more and more relevant in real-world applications, particularly when we lack information about boundary and initial conditions. In recent years, physics-informed neural networks (PINNs) have become a popular tool to solve PDEs for dynamical systems as the heat equation. PINNs can especially be used for unbounded problem settings to assimilate empirical data measurements with formal PDE models. However, PINNs often run into computational problems in high-dimensional settings as for time-dependent PDEs, as they require dense collocation points over the entire spatial-time computational domain. To tackle this challenge, we extend physics-informed boundary integral networks (PIBI-Nets) for the heat equation that only requires collocation points at the computational spatial-time domain boundary. PIBI-Net achieves comparable results as PINNs for the heat equation in unbounded domains, and even outperforms it in several practical settings. |