DM3L Inaugural Symposium 2024

We are live!

10.09. - 12.09.2024

Organized by: R. Furrer, F. Robmann, N. Trolese

Title: "Searching for Structure in Unfalsifiable Claims"
"While advances in automated fact-checking are critical in the fight against the spread of misinformation in social media, we argue that more attention is needed in the domain of unfalsifiable claims. In this talk, we outline some promising directions for identifying the prevailing narratives in shared content (image & text) and explore how the associated learned representations can be used to identify misinformation campaigns and sources of polarization."

Title: "Attention Models and Reasoning Self-Discovery"
"In this talk I will first present the standard attention-based operations that have been so successful for large language models. Then I will show preliminary results for the self- generation of non- NLP reasoning tasks that expend a basic set of simple hand-designed reasoning challenges. This approach provides a blueprint of a possible strategy to go beyond training on human- generated data or through reinforcement learning"

Title: "New faces of Bias in Online Platforms"
"The internet is fundamentally changing how we socialize, work, or gather information. The recent emergence of content serving services creates a new online ecosystem in which companies constantly compete for users' attention and use sophisticated user tracking and personalization methods to maximize their profit. My research investigates the potential downsides of the algorithms commonly used by online platforms. Since these algorithms learn from human data, they are bound to recreate biases that are present in the real world. In this talk, I will first present a measurement metholodogy developed to monitor personalization algorithms in the context of platforms such as Google Search or online stores. Second, I will talk about recommendation and rating systems in the context of employment related platforms such as job search sites, freelancing marketplaces and online professional communities. and their danger to reinforce gender inequalities."

Title:"Linking data and theory in network science using statistical inference"
"The number of network science applications across many different fields has been rapidly increasing. Surprinsingly, the development of theory and domain-specific applications often occur in isolation, risking an effective disconnect between theoretical and methodological advances and the way network science is employed in pratice. Here I will address the risk constructively, discussing good pratices to guarantee more successful applications and reproducible results. I will motivate designing statistically grounded methodologies to address challenges in network science. This approach allows one to explain observational data in terms of generative models, naturally deal with intrinsic uncertainties, and strengthen the link between theory and applications."

Title:"Conditional Flows for Interpretable Machine Learning"
"This talk is about the design of ML systems that are inherently transparent, usable, and interpretable. For this purpose, I will introduce three different strategies for enabling users to comprehend and reason about entire ML systems. These are: (i) the inclusion of physical constraints, (ii) the idea of building sparse models and (iii) the concept of human simulatability, defined as the extent to which a model can be simulated by a human within a reasonable timeframe. On the methodological side, the concept of conditional normalizing flows will define a foundation for all these different approaches."

Marcel Salathé (Salathé Lab, EPFL)

Title: "Science in the age of AI"
"In March 2024, OpenAI's CTO Mira Murati stated that GPT models could achieve PhD-level intelligence within the next two years. While such concrete timelines are notoriously difficult to predict, the rapid advancement of machine learning capabilities is undeniably going to transform the scientific landscape. This talk will explore the evolving relationship between AI and scientific research, drawing from personal experience, and work done is this area. The talk will cover two main areas: How are scientists and students using AI tools in their work today, and what potential ways should we except AI to change core scientific practices in the near future? I'll also discuss what these changes mean for science as a whole, including how we conduct research, teach new scientists, and collaborate within the scientific community."

Title: "Combining Machine Learning and Numerical Analysis for Partial Differential Equations"
"Partial differential equations enjoy widespread use as mathematical models of dynamics in space and time. Exploiting conncections between machine learning and numerical analysis for partial differential equations holds great potential for both areas. I will exemplify this in three cases: First I show the design of numerically consistent deep neural networks, which enable forecasting of chaotic dynamics from data. Second, I show how it enables noise-robust symbolic regression of partial differential equations from data with guaranteed physical consistency. Finally, I use the connection to explain and remedy training pathologies in physics-informed neural networks, extending their use to multi-scale and nonlinear problems and to Bayesian uncertainty quantifications over partial differential equation models. These developments present exciting opportunities in applications from spatial biology, which are dominated by nonlinear processes in space and time with often unknown physics. This is showcased in our work on understanding biological tissue morphogenesis as a self-organized mechano-chemical process."

Title: "Reconstruction and Synthesis of 3D Humans in 3D Scenes"
"Reconstructing and synthesizing 3D humans in 3D scenes is an important topic in computer vision and graphics. In this talk, I will present three lines of work. First, I will discuss the challenges and methods of reconstructing 3D humans from monocular videos, with a particular focus on ego-centric perspectives. Second, I will discuss our recent work on synthesizing natural human behaviours withtin 3D scenes. Last, I will present the application of our human motion synthesis models for generating synthetic datasets, thereby facilitating the training of more robut perception models for egocentric tasks."

Title: "Multimodal Data Integration: From Biomarkers to Mechanisms"
"An exciting oppurtunity at the intersection of the biomedical sciences and machine learning stems from the growning availability of large-scale multi-modal data (imaging-based and sequencing-based, observational and perturbational, at the single-cell level, tissue-level, and organism-level). Traditional representation learning methods, although often highly successful in predictive tasks, do not generally eludicate underlying causal mechanisms. I will present a statistical and computational framework for causal representation learning and its applications towards identifying novel disease biomarkers as well as inferring gene regulation in different disease contexts."