Session: Optimization as the Engine of Generative AI - II
Chair: Renyuan Xu
Cluster: Optimization for Emerging Technologies (LLMs, Quantum Computing, ...)
Talk 1: Designing and Optimizing Biological Molecules with Multimodal Stochastic Interpolant Generative Models
Speaker: Ge Liu
Abstract: The design and optimization of biological molecules, such as proteins and peptides, offers transformative potential in medicine, materials science, and synthetic biology. Traditional methods such as directed evolution often struggle to explore the vast and complex molecular landscape efficiently. In addition, molecule design problems inherently involve both discrete and continuous variables (e.g., protein sequences and 3D structures) and operate on non-Euclidean manifolds to model geometric information (e.g., rotational group SO(3)). Generative modeling has emerged as a powerful framework for biological molecule design. In this talk, I will present recent advances in SDE/ODE-based stochastic interpolant generative models, such as diffusion and flow-matching, that enabled precise and controllable generation of biological molecules across multiple modalities. First, I will introduce Statistical Flow Matching (SFM), a novel generative framework leveraging the Riemannian geometry of statistical manifolds that enables efficient generation of discrete data. SFM has demonstrated strong performance on biological sequence design (DNA, protein) and generalizable to text and image domains. Next, I will introduce OC-Flow, a theoretically grounded training-free optimal control framework for guiding flow-matching generative models on both Euclidean and non-Euclidean manifolds. By formulating generative sampling as an optimal control problem, OC-Flow enables effective guided sampling for solving a diverse set of inverse problem across computer vision, chemical molecule, and peptide design tasks, achieving controlled generation of molecules with optimized properties and energies. This talk will provide new perspectives on how stochastic interpolant generative models can bridge optimization, machine learning, and biomolecular engineering, paving the way for next-generation protein design.
Talk 2: Panel
Speaker: Ahmad Beirami
Abstract: Ahmad Beirami (Google DeepMind) and Renyuan Xu (Stanford University) will host a panel on the interactions between Optimization and Generative AI.
Talk 3: Panel
Speaker: Renyuan Xu
Abstract: Ahmad Beirami (Google DeepMind) and Renyuan Xu (Stanford University) will host a panel on the interactions between Optimization and Generative AI.
