一、报告题目：Flows in Probability Spaces for Image and Language Generation

二、报告人：Gabriele Steidl教授

三、时  间：2026年10月26日 9:00-10:30

四、地  点：A3-217

报告摘要：Generative flows, in particular flow matching provides state-of-the-art performance for sampling from (conditional) continuous distributions.The talk gives a brief introduction into the topic. In particular, we address recent research on learning the latent distribution and extend the approach to spherical flows for sampling from discrete distributions. Large language models (LLMs) based on autoregressive decoding are the prevailing approach to text generation. An alternative to sequential generation is to apply diffusion or flow-based generative modeling to produce the entire sequence at once. We study the problem of learning generative models for discrete distributions in a continuous embedding space. Whereas prior approaches typically operate on the Euclidean space or the probability simplex, we instead work on the n-sphere. There the von Mises-Fisher (vMF) distribution induces a natural noise process and admits a closed-form conditional score.Exploiting the radial symmetry of the vMF density, we reduce the continuity equation on the n-sphere to a scalar ODE in the cosine similarity,whose unique bounded solution determines the velocity of the spherical flow.

报告人简介：Gabriele Steidl is a Professor of Applied Mathematics at Technische Universität Berlin and an internationally renowned expert in mathematical imaging and applied mathematics. Her research focuses on computational harmonic analysis, convex optimization, inverse problems, optimal transport, and theory‑driven machine learning and generative modeling. She was elected SIAM Fellow in 2022 and serves as Editor‑in‑Chief of SIAM Journal on Imaging Sciences. In 2026, she was selected as a CNRS dFellow‑Ambassador. She devotes herself to solving practical problems in image‑signal processing and data analysis with rigorous mathematical theories. She has extensive teaching and research experience at multiple European universities and maintains broad international academic collaborations.

欢迎大家参加！联系人：喻高航