Unified framework for forecasting future states by harnessing diverse, multi-modal data streams.
By Kaouther MESSAOUD

Kaouther MESSAOUD will give a talk on “Unified framework for forecasting future states by harnessing diverse, multi-modal data streams”.

Abstract

In an era where dynamic, real-world environments demand rapid and reliable decision-making, predictive modeling stands at the forefront of advancing artificial intelligence. This presentation introduces a unified framework for forecasting future states by harnessing diverse, multi-modal data streams—ranging from 2D/3D sensor inputs to contextual cues—and by explicitly modeling both spatial and temporal interactions. By uncovering hidden structures and causal relationships within complex systems, our approach not only enhances generalization across heterogeneous domains but also generates multiple plausible outcomes to effectively capture inherent uncertainties. While the framework is broadly applicable, we highlight its practical utility in trajectory prediction for autonomous driving. In such settings, accurate forecasts of vehicle and pedestrian movements are critical for collision avoidance, path planning, and overall traffic safety. Our methodology integrates robust representation learning, adaptive prompt tuning for efficient domain adaptation, and security measures against adversarial manipulations, ensuring that predictive systems remain both flexible and resilient. This seminar will explore the core pillars of our approach—multi-modal input integration, interaction modeling, probabilistic outcome generation, parameter-efficient fine-tuning, and security enhancements—and demonstrate how these elements synergize to improve predictive performance in challenging, real-world scenarios.

Biography

Dr. Kaouther Messaoud is a scientist at École Polytechnique Fédérale de Lausanne (EPFL), specializing in deep learning, representation learning, and motion prediction. She earned her Ph.D. from INRIA and Sorbonne University, where she focused on attention-based trajectory prediction for autonomous driving. Her research explores ways to make predictive models more generalizable, efficient, and robust, particularly through state-of-the-art transformer-based architectures. She has collaborated with institutions such as University of California San Diego (UCSD) and Valeo AI, contributing to advancements in motion forecasting and AI-driven decision-making. Her work has been recognized at IEEE IV, ECCV, and ICLR, and she was honored with the George N. Saridis Best Transactions Paper Award (IEEE TIV).