This colloquium is intended to bring together Computer Science and Data Science researchers in the tri-state area (especially in NYC) and to foster collaboration. We welcome talks on any topic of interest to the CS community, including theory, algorithms, machine learning, and data science. If you are interested in attending in-person or online, or would like to give a talk, please contact Seminar Organizer, Jun Li at jun.li@qc.cuny.edu

Monday, 2/10/2025, 12:15pm – 1:30pm
Room: Science Building, C205
Speaker: Ziyi Huang, Bell Labs

Abstract: Artificial Intelligence (AI) has demonstrated immense potential to revolutionize healthcare, promising enhanced diagnostics, personalized treatments, and improved patient outcomes. However, the widespread adoption of AI in real-world clinical settings has been hindered, primarily due to concerns surrounding trustworthiness. In this talk, I will explore the concept of trustworthy AI in healthcare, examining its critical components and showcasing our research on enhancing AI trustworthiness for healthcare applications. I will introduce three specific areas of our work: robust tissue analysis, adaptive precision medicine, and machine learning safety analysis. My research on weakly supervised learning enables robust tissue analysis from imperfectly annotated datasets, substantially reducing the workload associated with data collection and annotation. Leveraging reinforcement learning techniques, I have developed models that can potentially provide tailored decision making, enabling adaptive, optimal, and personalized treatment strategies with statistical guarantees. Furthermore, my work on safety analysis centers on quantifying model uncertainty, providing practical guidance to enhance the reliability of deep neural networks. Finally, I will share my broader research vision on the trustworthy smart health system and its potential applications across diverse healthcare domains.

Dr. Ziyi Huang is currently a machine learning researcher at Bell Labs and an adjunct faculty at Stevens Institute of Technology. Prior to these roles, she got her Ph.D. degree in Electrical Engineering from Columbia University, her master’s degree from the University of Michigan, and her bachelor’s degree from the University of Science and Technology of China (USTC). Her research primarily focuses on trustworthy AI for health, computational biology, and image-guided therapy. Her vision is to empower machines to deliver trustworthy disease analysis and reliable treatment strategies across populations. Dr. Huang’s work has been accepted at top venues, including IEEE-JBHI (IF 7.7), NeurIPS, AISTATS, and MICCAI. Her contribution has been recognized by selective fellowships/grants with over $283K awarded, and she was selected as Machine Learning and Systems Rising Star in 2024.

Monday, 2/24/2025, 12:15pm – 1:30pm
Room: Science Building, C207
Speaker: Hoa Wang, Stevens Institute of Technology

Abstract: Deep reinforcement learning (DRL) algorithms are widely applicable in many different areas, such as scientific simulations, robotics, autonomous driving, and large language model (LLM) development. However, DRL training is computationally expensive, requiring numerous trial-and-errors and consuming substantial computing resources and time. From an algorithmic perspective, the stochastic nature of environment dynamics can cause some actors to complete episodes sooner, resulting in idle periods while waiting for other actors to finish. From a systems perspective, actors remain idle during the policy update by the learner, significantly wasting computing resources and amplifying training costs.

In this talk, I will introduce our recent studies published by AAAI’24 and accepted by SC’24. The SC’24 paper has received the Best Student Paper Nomination. First, I will address the fundamental challenges in large-scale distributed DRL training. I will then delve into our recent studies on distributed DRL training using serverless computing. Serverless computing, also known as Function-as-a-Service (FaaS), is a cloud computing model that employs lightweight containers as execution units. The instant execution and auto-scaling capabilities of serverless computing naturally meet the highly dynamic resource demands of DRL training. Finally, I will discuss our ongoing research and future outlook for serverless computing architectures and LLM inference.

Monday, 3/17/2025, 12:15pm – 1:30pm
Room: Science Building, C205
Speaker: Nairen Cao, New York University

Abstract: Correlation Clustering is one of the most extensively studied graph clustering problems. The input is a complete graph where each edge is labeled either “+” or “−,” and the objective is to partition the vertex set into an arbitrary number of clusters to minimize the sum of (i) the number of “+” edges spanning different clusters and (ii) the number of “−” edges within the same cluster. Until recently, the best known polynomial-time approximation ratio was 2.06, nearly matching the integrality gap of 2 for the standard LP relaxation. Subsequent work leveraging the Sherali-Adams hierarchy improved this bound to 1.73.

To establish a simple yet powerful framework for Correlation Clustering—akin to those used in typical approximation algorithms—we introduce the Clustering LP, a strong linear program that we conjecture closely captures the approximability of the problem. We demonstrate the strength of the Clustering LP by presenting a simple rounding algorithm that achieves a 1.437-approximation, surpassing previous results. Furthermore, we show that the Clustering LP can be solved in sublinear time O(n) or in O(1) rounds in the MPC model, or only \tilde{O}(n) space streaming model using, where n is the number of vertices in the graph. These results establish the state-of-the-art in terms of both approximation ratio and running time in several models.

This work is a joint collaboration with Vincent Cohen-Addad, Shi Li, Euiwoong Lee, David Rasmussen Lolck, Alantha Newman, Mikkel Thorup, Lukas Vogl, Shuyi Yan, and Hanwen Zhang.

Wednesday, 3/26/2025, 12:15pm – 1:30pm
Room: Science Building, C205
Speaker: Jian Li, Stony Brook University

Abstract: Next-generation (NextG) networked systems, supporting applications such as 5G/6G, autonomous driving, and augmented/virtual reality, will be increasingly data-driven and large-scale. This shift places enormous demands on both traditional wireless communication networks and emerging architectures like edge and cloud computing systems, where classical algorithmic guarantees may no longer hold, and rigorous performance evaluation becomes increasingly challenging. While various data-driven techniques have recently been applied to networking and distributed systems, purely data-driven solutions face challenges such as the curse of dimensionality, limited expressiveness, and generalization issues. In this talk, I will present a data-driven structured reinforcement learning framework that addresses these challenges by judiciously combining model-based and model-free learning techniques. This novel framework leverages the inherent structure encoded in classical models, enabling the design of reinforcement learning architectures with strong theoretical performance guarantees (e.g., order-of-optimal regret) and low computational complexity, making them practical for real-world implementation. Additionally, I will share my vision for NextG networked systems, expanding the scope of structured learning to enable more tractable system design. This includes, but is not limited to, applications in Large Language Models (LLMs) and other emerging AI-driven technologies.

Bio: Jian Li is an Assistant Professor of Data Science at Stony Brook University. He received his Ph.D. in Computer Engineering from Texas A&M University in 2016, and his B.E. from Shanghai Jiao Tong University in 2012. His current research interests lie at the intersection of algorithms for reinforcement learning and multi-armed bandits, decentralized/federated learning, and stochastic optimization and control, with applications to next-generation networked systems. He is a recipient of the NSF CAREER Award (2024) and NSF CISE CRII Award (2021).

Wednesday, 4/2/2025, 12:15pm – 1:30pm
Room: Science Building, C205
Speaker: Nikhil Garg, Cornell Tech

Abstract: Modern societal and governmental decisions — what gets built and maintained, who gets access to what school, where people interview and work — are driven by computational systems. While much of the focus is on learning from data, many of the real-world challenges appear before and after the learning algorithm: do people equitably engage with the system; can we collect accurate, unbiased, up-to-date data; and then how should we use individually noisy predictions to make efficient, equitable societal decisions, given resource constraints and population (market and strategic) effects? In this talk, I’ll discuss our theoretical, empirical, and deployment work in tackling these challenges, in collaboration with NYC government agencies, a university admissions team, a platform to help discharge patients to long-term care facilities, and deploying feed ranking algorithms on Bluesky. I’ll focus on recommendation-matching systems: (1) ongoing work on the NYC High School match, to understand disparities in ranking behavior and provide recommendations to applicants; (2) the role of algorithmic monoculture and noise in matching markets.

Thursday, 4/10/2025, 12:15pm – 1:30pm
Room: Science Building, C203
Speaker: Peter Brass, City College of CUNY

Abstract: In this presentation I will describe how NSF works, what a program director does, the process by which proposals are decided. Then I will give advice on how to find the right program, how to write proposals, how to interact with the NSF, and what to do if you get funded. The advice should apply similar to computer science, math, and closely related subjects.

Bio: Peter Brass studied mathematics and computer science at the Technical University Braunschweig in Germany, where he received his PhD in 1992, followed by a postdoc period at the University of Greifswald 1993-1997 and a Heisenberg Fellowship 1998-2002 at the Free University Berlin, where his research moved from Discrete Geometry to Algorithms. Since 2002 he is professor of computer science at CCNY, and 2021-2025 he served as program director at the NSF, in the Algorithmic Foundations (Lead), Robotics, REU Sites, and several other programs. He wrote three books, “Research Problems in Discrete Geometry” (with J. Pach, W. Moser), “Advanced Data Structures”, and “History of Engineering at the City College of New York”, as well as about fifty journal papers on problems in geometry and algorithms.

Monday, 4/28/2025, 12:15pm – 1:30pm
Room: Science Building, C207
Speaker: Adarsh Srinvasan, Rutgers University

Abstract: