Abstracts/Bios

Sanghamitra Dutta

Title: Information-theoretic Methods in Explainability for High-Stakes Applications

Abstract: How do we ensure that the machine learning algorithms in high-stakes applications are fair, explainable, and lawful? Towards addressing this urgent question, this talk provides some strategies deep-rooted in information-theoretic methods. In the first part of the talk, I will discuss an emerging problem in explainability, also called robust counterfactual explanations: how do we guide a rejected applicant to receive a favourable model outcome while also being robust to model multiplicity (Rashomon Effect)?  We propose strategies to provide counterfactual explanations that remain robust under model changes with high probability. In the second part of the talk, I will introduce an information-theoretic tool called Partial Information Decomposition and discuss its role in fairness and explainability problems.

Bio: Sanghamitra Dutta is an assistant professor in the Department of Electrical and Computer Engineering at the University of Maryland College Park since Fall 2022. She is also affiliated with the Center for Machine Learning (CML) at UMIACS. Prior to joining UMD, she was a senior research associate at JPMorgan Chase AI Research New York in the Explainable AI Centre of Excellence (XAI CoE). She received her Ph.D. and Masters's from Carnegie Mellon University and B. Tech. from IIT Kharagpur, all in Electrical and Computer Engineering.

Her research interests broadly revolve around reliable and trustworthy machine learning where she brings in novel foundational perspectives deep-rooted in information theory, statistics, causality, and optimization. In her prior work, she has also examined problems in reliable computing for large-scale distributed machine learning, using tools from coding theory (an emerging area called “coded computing”).

She is a recipient of the 2024 NSF CAREER Award, 2023 JP Morgan Faculty Award, 2023 Northrop Grumman Seed Grant, 2022 Simons Institute Fellowship for Causality, 2021 AG Milnes Outstanding Thesis Award from CMU and 2019 K&L Gates Presidential Fellowship in Ethics and Computational Technologies. She has also pursued summer research internships at IBM Research and Dataminr.

Catuscia Palamidessi

Title:  Information Structures for Privacy and Fairness

Abstract: The increasingly pervasive use of big data and machine learning raises various ethical issues, particularly privacy and fairness. In this talk, I will discuss some frameworks to understand and mitigate the issues, focusing on iterative methods coming from information theory and statistics. In privacy protection, differential privacy (DP) and its variants are the most successful approaches to date. One of the fundamental issues of DP is how to reconcile the loss of information that it implies with the need to preserve the utility of the data. In this regard, a useful tool to recover utility is the Iterative Bayesian Update (IBU), an instance of the famous Expectation-Maximization method from Statistics. I will show that the IBU, combined with the metric version of DP, outperforms the state-of-the art, which is based on algebraic methods combined with the Randomized Response mechanism, widely adopted by the Big Tech industry (Google, Apple, Amazon, ...). Furthermore, I will discuss a surprising duality between the IBU and one of the methods used to enhance metric DP, which is the Blahut-Arimoto algorithm from Rate-Distortion Theory. Finally, I will discuss the issue of biased decisions in machine learning and will show that the IBU can be applied also in this domain to ensure a fairer treatment of disadvantaged groups.

Bio: Catuscia Palamidessi is Director of Research at Inria Saclay (since 2002), where she leads the team COMETE. She has been a Full Professor at the University of Genova, Italy (1994-1997) and at Penn State University, USA (1998-2002). 

Palamidessi's research interests include Privacy, Machine Learning, Fairness, Secure Information Flow, Formal Methods, and Concurrency. Her past achievements include the proof of expressiveness gaps between various concurrent calculi, and the development of a probabilistic version of the asynchronous pi-calculus. More recently, she has contributed to establishing the foundations of probabilistic secure information flow, she has proposed a metric extension of differential privacy, and geo-indistinguishability, an approach to location privacy, which received the ToT award at ACM CCS 2023. 

Palamidessi is leading various scientific projects, including an ERC advanced grant to conduct research on Privacy and Machine Learning. In recognition of her research achievements, in 2022 she received the Grand Prix of the French Academy of Science. 

Palamidessi is on the Editorial Board of several journals, including the IEEE Transactions in Dependable and Secure Computing, the ACM Transactions on Privacy and Security, the Journal of Computer Security, Mathematical Structures in Computer Science, and Acta Informatica. 

Mario Diaz

Title: On the Differential Privacy Guarantees of Iterative Training Algorithms

Abstract: Iterative algorithms are extensively employed in the private training of machine learning models, e.g., Differentially Private Stochastic Gradient Descent. In this talk, we present effective bounds for the privacy loss of general iterative algorithms, without relying on convexity or smoothness assumptions regarding the loss function. Our methodology relies on an analysis of the hockey-stick divergence between coupled non-Markovian stochastic processes.

Bio: Mario Diaz received the B.Eng. degree in electrical engineering from the Universidad de Guadalajara, Guadalajara, Mexico, in 2011, the M.Sc. degree in probability and statistics from the Centro de Investigación en Matemáticas, Guanajuato, Mexico, in 2013, and the Ph.D. degree in mathematics and statistics from Queen’s University, Kingston, Canada, in 2017. He is currently a Research Associate with the Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas (IIMAS), Universidad Nacional Autónoma de México, Mexico City, Mexico. Prior to this, he was a Post-Doctoral Scholar with Arizona State University, the Centro de Investigación en Matemáticas, and Harvard University. His research interests include the mathematical and statistical foundations of information privacy, theoretical machine learning, and random matrix theory.

Kush Varshney

Title: The AI Alignment Problem, Solutions, and Limits

Abstract: With the recent rise of large language models (LLMs) and generative AI, we are treating interactions - both helpful and harmful - as behaviors that should be aligned with some notion of human values. In this talk, I will first describe LLMs and their new and amplified risks and harms, such as hallucination, prompt injection, information leakage, copyright infringement, bullying, and gaslighting. Then I will overview the alignment problem and different formulations, including different kinds, horizons, and audiences. I will discuss information-theoretic limits to alignment and present some constructive approaches. I will conclude with some thoughts on value pluralism, moral philosophy, and the decoloniality of knowledge.

Bio: Kush R. Varshney was born in Syracuse, New York in 1982. He received the B.S. degree (magna cum laude) in electrical and computer engineering with honors from Cornell University, Ithaca, New York, in 2004. He received the S.M. degree in 2006 and the Ph.D. degree in 2010, both in electrical engineering and computer science at the Massachusetts Institute of Technology (MIT), Cambridge. While at MIT, he was a National Science Foundation Graduate Research Fellow.

Dr. Varshney is an IBM Fellow, based at the Thomas J. Watson Research Center, Yorktown Heights, NY, where he heads the Human-Centered Trustworthy AI team. He was a visiting scientist at IBM Research - Africa, Nairobi, Kenya in 2019. He applies data science and predictive analytics to human capital management, healthcare, olfaction, computational creativity, public affairs, international development, and algorithmic fairness, which has led to the Extraordinary IBM Research Technical Accomplishment for contributions to workforce innovation and enterprise transformation, IBM Corporate Technical Awards for Trustworthy AI and for AI-Powered Employee Journey, and the IEEE Signal Processing Society’s 2023 Industrial Innovation Award.

He and his team created several well-known open-source toolkits, including AI Fairness 360, AI Explainability 360, Uncertainty Quantification 360, and AI FactSheets 360. AI Fairness 360 has been recognized by the Harvard Kennedy School's Belfer Center as a tech spotlight runner-up and by the Falling Walls Science Symposium as a winning science and innovation management breakthrough.

He conducts academic research on the theory and methods of trustworthy machine learning. His work has been recognized through paper awards at the Fusion 2009, SOLI 2013, KDD 2014, and SDM 2015 conferences and the 2019 Computing Community Consortium / Schmidt Futures Computer Science for Social Good White Paper Competition. He independently-published a book entitled 'Trustworthy Machine Learning' in 2022, available at http://www.trustworthymachinelearning.com. He is a fellow of the IEEE.

Haewon Jeong

Title: Emerging safety issues in generative models

Abstract: The Internet is bound to be filled with AI-generated content, including AI-generated images and videos. While they look photorealistic, are they also safe to use for training new generative models? If we keep using synthetic images to finetune diffusion models, is there going to be a catastrophic failure? We will discuss an interesting “mode collapse” phenomenon in the chain of finetuning diffusion models with synthetic images. We then discuss how we should generate “reusable” images and propose Reusable Diffusion Finetuning (ReDiFine) that can achieve a better fidelity-reusability tradeoff. I will also briefly talk about fairness in facial phenotype reconstruction when we use generative models for image compression.

Bio: Jeong received a Ph.D. in Electrical and Computer Engineering at Carnegie Mellon University. Her thesis established important foundations on how we can apply coding theory to building reliable large-scale computing systems. Before joining UCSB, she was a postdoctoral fellow at Harvard University, where she explored reliability in a different sense: how to build a machine learning system humans can trust and rely on. In particular, she investigated how machine learning systems can discriminate against students in education-related applications, and how we can build more fair machine learning algorithms. She is a recipient of the NSF CAREER Award, JP Morgan Faculty Award, and named as a 2024 Hellman fellow.

Ananda Theertha Suresh

Title: Improved Private Mean Estimation

Abstract: Mean estimation is a fundamental problem in statistics and is often used as a subroutine in practical applications. In many scenarios, the underlying data is private and hence differential private algorithms are used to estimate the mean. The standard folklore private approach involves adding Laplace noise to both the sum and the dataset size, and estimate the mean as their ratio.  In this talk, we demonstrate that this folklore well-used method is suboptimal and develop techniques that improve upon it.  

We first propose a new algorithm and show that it is min-max optimal, that it has the correct constant in the leading term of the mean squared error. In addition, we demonstrate empirically that our proposed algorithm yields a factor of two improvement in mean squared error over methods often used in practice.  We then propose a new definition of instance-optimality for differentially private estimation and show that this definition is significantly stronger than those proposed in prior work. We develop an instance-optimal mean estimation algorithm under this notion of instance-optimality.  Finally, we demonstrate the applications of this new method, by using it in machine learning.  We develop a new algorithm called DPSGD-F that combines our mean estimation method with the standard DPSGD  algorithm. For linear classification tasks, we theoretically show that it incurs smaller error than DPSGD and demonstrate its practicality on image classification benchmarks.

Bio: Ananda Theertha Suresh is a staff research scientist at Google Research, New York. He received his PhD from University of California San Diego, where he was advised by Prof. Alon Orlitsky. His research lies in the intersection of information theory and machine learning. His recent research is primarily centered on developing algorithms and theoretical formulations for addressing efficiency, resource, and privacy considerations for machine learning systems in practice.  He is a recipient of the 2017 Paul Baran Maroni Young Scholar award and a co-recipient of best paper awards at NeurIPS 2015, ALT 2020, CCS 2021.