Frontiers in Electrical Engineering

Anastasios Angelopoulos

Title: Statistical Foundations of Trustworthy AI Engineering

Image Lightbox

Download Full Image

Abstract: My research has been almost entirely devoted to a single question:

How can we build trustworthy systems from untrustworthy AI algorithms?

Answering this question is difficult because modern AI models can be wrong in unpredictable ways. From data, these models learn biases, spurious associations, and imperfect world-models that are difficult to debug due to their statistical nature. But to use AI in critical applications---from legal and financial institutions to power plants to hospitals, where safety, and lives are at stake---we need trust. Part of what holds us back is a lack of formally grounded but practical statistical methodology for ensuring that we are able to use AI reliably, even when the underlying model may have flaws.

The talk will have two halves. In the first half, I will discuss conformal risk control, a statistical framework for reliable decision-making using black-box models.  In the second half, I will discuss AI evaluations for aligning AI with human preferences and safety. I will focus both on the foundational statistical methodology underlying these techniques and also the large-scale deployments that have resulted, and the opportunities for future research that arise.

Bio: Anastasios Nikolas Angelopoulos is a sixth-year Ph.D. student at the University of California, Berkeley. Previously, he obtained a B.S. in Electrical Engineering at Stanford University. His research concerns statistical infrastructure for reliable and safe deployment of AI, including conformal prediction, prediction-powered inference, and the development of Chatbot Arena, an open-source platform measuring the alignment of AI to human preferences.

Jason Ma

Title: Internet Supervision for Robot Learning

Abstract: The availability of internet-scale data has led to impressive large-scale AI models in various domains. For learning robot skills, despite recent efforts in crowd-sourcing robot data, robot-specific datasets remain orders of magnitude smaller than datasets used in vision or language foundation models. Rather than solely focusing on scaling robot data, my research takes the alternative path of directly using available internet data and models as supervision for robots -- in particular, learning general feedback models for robot actions. Feedback can be relatively agnostic to robot embodiments, applicable to various policy learning algorithms, and as I will show, can be learned even from exclusively non-robot data. I will present two complementary approaches in this talk. First, I will present a novel reinforcement learning algorithm that can directly use in-the-wild human videos to learn value functions, producing zero-shot dense rewards for manipulation tasks specified in images and texts. Second, I will demonstrate how grounding large language models code search with simulator feedback enables automated reward design and beyond for sim-to-real transfer of complex robot skills, such as a quadruped robot dog balancing on a yoga ball. 

Bio: Jason Ma is a fifth-year PhD student at the University of Pennsylvania. His research interests span robot learning, reinforcement learning, and deep learning. His work has received Best Paper Finalist at ICRA 2024, Top 10 NVIDIA Research Projects of 2023, and covered by popular media such as the Economist, Fox, Yahoo, and TechCrunch. Jason is supported by Apple Scholar in AI/ML PhD Fellowship as well as OpenAI Superalignment Fellowship.

Xinyi Chen

Title: Principled Algorithms for Efficient Machine Learning: A Dynamical Systems View

Image Lightbox

Download Full Image

Abstract: Machine learning has driven many technological breakthroughs, yet training neural networks has become increasingly expensive. In this talk, I will present my research on developing theoretically-founded methods to improve the efficiency of machine learning, with a focus on optimization. Finding the best training algorithm and tuning its parameters is a costly but often necessary part of the training process. However, finding the best algorithm for particular optimization instances is a non-convex problem that is challenging for theory and practice. We will begin by describing meta-optimization, a framework for learning the best optimizer from problem instances. We will then discuss how meta-optimization can be formulated as a feedback control problem, and how recent advances in online control leads to provable methods for meta-optimization. We will conclude with how foundational tools in control theory can advance efficient architecture design, highlighting the expanding intersection between optimization, control theory, and machine learning.

Bio: Xinyi Chen is a PhD candidate in the Department of Computer Science at Princeton University advised by Prof. Elad Hazan. She is also a research scientist at Google DeepMind. Her research is at the intersection of machine learning, optimization, and dynamical systems, in particular developing provably efficient methods for sequential decision-making and control. Previously, she obtained her undergraduate degree from Princeton in Mathematics, where she received the Middleton Miller Prize. Among other honors, she has received the NSF Graduate Research Fellowship and the Siebel Scholarship, and has been recognized by EECS Rising Stars at UC Berkeley and Rising Stars in Data Science at the University of Chicago. 

Anuran Makur

Title: Information and Learning Theory in Adverse Settings: Tournament Rankings and Permutation Channels

Image Lightbox

Download Full Image

Abstract: As new techniques for processing data are perpetually emerging, it has become ever more important to develop foundational concepts in machine learning and information theory. In this vein, my research develops both theory and algorithms for various problems of inference, learning, computation, and communication. In this talk, I will discuss two such threads of work on information and learning theory in very noisy contexts. First, I will present the problem of learning rankings of agents in a tournament from pairwise comparison data. This problem naturally emerges in settings like sports, financial markets, and even reinforcement learning from human feedback. Perhaps the most popular approach to learning rankings involves assuming a Bradley-Terry-Luce (BTL) model and estimating its deterministic parameters, which represent unknown skill levels of agents. I will fundamentally develop this narrative in two ways. On the one hand, I will illustrate that if we assume that BTL parameters are drawn from an unknown probability distribution, we can also estimate the entire skill distribution using a surprisingly simple algorithm in a minimax sense. This gives us the ability to calculate functionals of the skill distribution, such as entropy, to define overall measures of skill in a tournament. On the other hand, I will present the first rigorous minimax hypothesis test to determine whether data actually obeys a BTL model in the first place, and demonstrate tight bounds on its critical threshold. Finally, I will also delineate the interplay between these results and ideas from non-parametric statistics, information theory, and the theory of Markov chains. Second, inspired by applications in DNA storage systems and communication in packet networks, I will introduce and analyze permutation channel and network models where codewords experience permutations. Specifically, a permutation channel consists of a discrete memoryless channel followed by a random permutation that reorders the output codeword of the channel. I will formalize an alternative notion of "rate" of information transmission as well as the associated fundamental limit of permutation channel capacity for such models. Then, I will discuss coding techniques to characterize and achieve capacities and capacity regions of various classes of permutation channels and networks, respectively, such as point-to-point and adder multiple-access channels. These results will illustrate an emerging theory of permutation channels and networks..

Bio: Anuran Makur is an Assistant Professor with the Department of Computer Science and the Elmore Family School of Electrical and Computer Engineering, Purdue University. He received the B.S. degree with highest honors from the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley (UC Berkeley) in 2013, and the S.M. and Sc.D. degrees from the Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology (MIT) in 2015 and 2019, respectively. From 2019 to 2021, he was a postdoctoral researcher at the Laboratory for Information and Decision Systems and the Institute for Data, Systems, and Society, MIT. His research interests are in the theory of machine learning, information theory, and applied probability. He was a recipient of the Arthur M. Hopkin Award from UC Berkeley in 2013, the Jacobs Presidential Fellowship from MIT in 2013, the Ernst A. Guillemin Master's Thesis Award from MIT in 2015, the Jin Au Kong Doctoral Thesis Award from MIT in 2020, the Thomas M. Cover Dissertation Award from the IEEE Information Theory Society in 2021, and the NSF CAREER Award in 2023.