[Seminar] MLDS Unit Seminar 2024-7 by Mr. Rahul Steiger (ETH Zurich), Mr. Thibault De Surrel (Université Paris-Dauphine) at C210

Speaker 1: Mr. Rahul Steiger (ETH Zurich)

Title: In-Context Predictions on ICU Time-Series

Abstract: The increase of patient data in intensive care units (ICU), coupled with the the lag in medical staffing, introduces an urgent need for scalable and efficient predictive decision support systems. Despite significant improvements in predictive modeling techniques and the abundance of patient data, effectively leveraging ICU data remains a formidable challenge. Traditional modeling approaches often encounter difficulties in adapting to varying treatment protocols and patient demographics. Inspired by the recent advancements in computer vision and natural language processing, we introduce a novel framework based on self-supervised learning (SSL) and a non-parametric, task-agnostic prediction mechanism. This predictive approach seamlessly integrates domain-shifted and patient history data without necessitating model retraining. Such a framework not only holds the potential to enhance predictive accuracies but also to enable personalized care interventions within the ICU setting. 
 

Speaker 2: Mr. Thibault De Surrel (Université Paris-Dauphine)

Title: How Riemannian geometry can help us build better Brain Computer Interfaces

Abstract: The goal of Brain-Computer Interfaces (BCI) is to translate brain electrical signals into actionable commands. We achieve this by using Electroencephalograms (EEGs), which measure the electrical activity of various brain regions. These EEG signals are transformed into covariance matrices, providing insights into the connectivity between different brain areas. Covariance matrices possess a unique structure as they are Symmetric Positive Definite (SPD) and thus live on a Riemannian manifold. In this talk, I will introduce the Riemannian geometry of SPD matrices and present two of my recent projects.
The first project involves adapting t-SNE to a Riemannian framework, enabling the visualization of high-dimensional SPD matrices in a 3D space while preserving the curvature of the high-dimensional manifold. This approach prevents distortions that occur when flattening the representation in a Euclidean context. We successfully utilized this method to visualize BCI datasets, offering more accurate and insightful representations.
The second project focuses on developing a Gaussian distribution on the manifold of SPD matrices. Our objective is to model a set of SPD matrices with a probability distribution that inherently accounts for Riemannian geometry. We accomplish this by mapping a Euclidean multivariate Gaussian distribution onto the manifold of SPD matrices. I will present preliminary results of this distribution and share experiments on parameter estimation using a maximum likelihood estimator.