Fwd: Monday seminar by Prof. Kai Wang (Georgia tech)

[Shaddin Dughmi]

This might be of interest to many of you. Kai  started as a PhD student at USC. 

Shaddin Dughmi

Associate Professor

Computer Science Department

University of Southern California (USC)
Email: sha...@usc.edu

Web: https://viterbi-web.usc.edu/~shaddin/

---------- Forwarded message ---------
From: Bistra Dilkina <dil...@usc.edu>
Date: Sun, May 18, 2025, 11:13 AM
Subject: Monday seminar by Prof. Kai Wang (Georgia tech)
To: David Kempe <david....@gmail.com>, Shaddin Dughmi <sha...@usc.edu>

Please share with the theory group faculty and students 

Location: GCS 402C

Time: 11-Noon followed by pizza

Title: A Fully First-Order Method for Decision-Focused Learning

Abstract: Decision-focused learning (DFL) aims to train predictive models by optimizing downstream decision quality rather than traditional predictive accuracy. While recent algorithms have integrated optimization directly into the training pipeline to achieve DFL, they often rely on second-order derivatives or surrogate losses, limiting their scalability and applicability in large-scale or constrained settings. In this talk, I will introduce a fully first-order method for decision-focused learning, grounded in recent advances in bilevel optimization. Our approach enables efficient end-to-end training using only first-order gradients by reformulating the decision problem with a smooth penalty, allowing gradients to be backpropagated through the optimization layer without requiring differentiation through exact solvers. I will discuss both the theoretical convergence guarantees of our method and initial empirical results demonstrating strong performance on large-scale problems. This work advances the scalability and practicality of decision-focused learning by eliminating key computational bottlenecks.

Bio: Kai Wang is an assistant professor in the school of Computational Science and Engineering at Georgia Institute of Technology. He received his Ph.D. in Computer Science at Harvard University. His research interests include AI for social impact, machine learning, optimization, and multi-agent systems, with a focus on applications in health and environmental sustainability. One of Kai's key contributions is the first real-world deployment of decision-focused learning to provide intervention recommendations in maternal health in collaboration with an Indian non-profit, where his algorithm has been deployed and is currently used by more than 350,000 people with a 31% improvement in health information engagement. Kai’s work has been recognized with the Schmidt Science AI2050 Early Career Fellowship, Siebel Scholars, and the best paper runner-up award at AAAI.

Bistra Dilkina
Dr. Allen and Charlotte Ginsburg Early Career Chair in Computer Science

Associate Professor of Computer Science and of Industrial & Systems Engineering 

Co-Director   |  CAIS Center for Artificial Intelligence in Society  |  www.cais.usc.edu

Co-Director  |  ORAI Interdisciplinary PhD Program on Artificial Intelligence and Operations Research  |  https://sites.google.com/usc.edu/usc-nrt-orai/

University of Southern California | Viterbi School of Engineering

University Park Campus | Office GCS 405H

dil...@usc.edu

https://viterbi.usc.edu/directory/faculty/Dilkina/Bistra