Linearized Augmented Lagrangian Methods for Nonconvex Functional Constrained Optimization



    Department of Systems Engineering and Engineering Management

                       The Chinese University of Hong Kong


Date: Friday, April 14, 4:30 pm – 5:30 pm

Venue: ERB 513, The Chinese University of Hong Kong

Title: Linearized Augmented Lagrangian Methods for Nonconvex Functional Constrained Optimization

Speaker: Dr. Songtao Lu, IBM Research



In this talk, I will discuss my recent work on designing inearized augmented Lagrangian methods for non-convex optimization. Specifically, I focus on solving non-convex functional constrained optimization problems, encompassing many resource-limited problems. While there have been existing methods for solving this class of problems, they typically involve double-loop or triple-loop algorithms that require oracles to solve subproblems with high accuracy by tuning multiple hyperparameters at each iteration.

To address this issue, I will introduce an augmented Lagrangian-based single-loop algorithm that can solve this class of problems without sacrificing iteration complexity while achieving faster numerical convergence compared to state-of-the-art methods used for practical machine learning problems such as multi-class Neyman-Pearson classification, constrained Markov decision processes, and neural net training with budget constraints. Next, I will showcase the extension of this method for solving bi-level functional constrained optimization problems, with applications to meta-causal structure discovery. Finally, I will conclude the talk by highlighting several promising research directions in non-convex functional constrained (bi-level) optimization.


Songtao Lu is currently a senior research scientist with the mathematics and theoretical computer science group at the IBM Thomas J. Watson Research Center, Yorktown Heights. He obtained his doctoral degree in electrical engineering from Iowa State University in 2018. He was a post-doctoral associate with the department of electrical and computer engineering at the University of Minnesota Twin Cities from 2018 to 2019, and an AI resident at the Thomas J. Watson Research Center from 2019 to 2020. Dr. Lu is a recipient of the best paper runner-up award of UAI (2022), the outstanding paper award of NeurIPS federated learning workshop (2022), and the IBM research accomplishment award (2021). His recent works have been published at multiple top-tier AI and machine learning conferences, including ICML, NeurIPS, AAAI, ICLR, UAI, IJCAI, AISTATS, etc. His primary research interests lie in machine learning, optimization, AI, and data science. 

Everyone is welcome to attend the talk!

SEEM-5202 Website:


Friday, April 14, 2023 - 16:30 to 17:30