[Seminar] MLDS Unit Seminar 2025-4 by Mr. Rémi Surat (Université Paris Cité, ENSAE Paris), Ms. Klea Ziu (MBZUAI)

Speaker 1: Mr. Rémi Surat (Université Paris Cité, ENSAE Paris) 

Title: Flow-based generative models: how to train them efficiently using flow matching and optimal transport

Abstract: This talk aims to present generative models based on flows, which are currently state-of-the-art methods in generative AI for continuous spaces. We will begin with a brief overview of energy-based models, one of the earliest approaches to model probability distributions, which provide the basic ideas for understanding the difficulties of data generation. We will then focus on flow-based models and their efficient training through flow matching, while also explaining their connections to the well-known diffusion models. Particular emphasis will be placed on how optimal transport can serve as a framework for defining dynamic flows in probability space. Finally, we will discuss further links with Schrödinger bridges, highlighting the unifying perspective they bring to modern generative modeling.

Speaker 2: Ms. Klea Ziu (Mohamed bin Zayed University of Artificial Intelligence: MBZUAI, UAE) 

Title: ψDAG: Projected Stochastic Approximation Iteration for Linear DAG Structure Learning

Abstract: Learning the structure of Directed Acyclic Graphs (DAGs) presents a significant challenge due to the vast combinatorial search space of possible graphs, which scales exponentially with the number of nodes. Recent advancements have redefined this problem as a continuous optimization task by incorporating differentiable acyclicity constraints. These methods commonly rely on algebraic characterizations of DAGs, such as matrix exponentials, to enable the use of gradient-based optimization techniques. Despite these innovations, existing methods often face optimization difficulties due to the highly non-convex nature of DAG constraints and the per-iteration computational complexity. In this work, we present a novel framework for learning DAGs, employing a Stochastic Approximation approach integrated with Stochastic Gradient Descent (SGD)-based optimization techniques. Our framework introduces new projection methods tailored to efficiently enforce DAG constraints, ensuring that the algorithm converges to a feasible local minimum. With its low iteration complexity, the proposed method is well-suited for handling large-scale problems with improved computational efficiency. We demonstrate the effectiveness and scalability of our framework through comprehensive experiments, which confirm its superior performance across various settings.