CDS Spring 2023 Colloquium
Recent Colloquium Talks
Tuesday, April 18, 2023
Towards a Statistical Foundation for Reinforcement Learning
Andrea Zanette, Postdoctoral Scholar in the Department of Electrical Engineering and Computer Sciences at the University of California at Berkeley
Time: 1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: In recent years, reinforcement learning algorithms have achieved a number of headline-grabbing empirical successes on various complex tasks. However, applying the reinforcement learning paradigm to new problems remains highly challenging. In many cases, the existing algorithms need to be modified, and new ones may have to be developed to solve the problem at hand. In order to do so effectively, we must gain some understanding about the foundations of reinforcement learning. In this talk, Zanette will present some recent results of his research towards this goal. He will first present an algorithm that can exploit the domain structure to learn much faster on easier problems, while retaining state-of-the art worst-case guarantees on pathologically hard ones. Then he will discuss a fundamental information-theoretic lower-bound, which establishes that reinforcement learning can be exponentially harder than supervised learning even when simple linear predictors are implemented. Finally, Zanette will discuss a statistically optimal algorithm to learn from historical data.
Andrea Zanette is a postdoctoral scholar in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley, supported by a fellowship from the Foundation of Data Science Institute. He completed his PhD (2017-2021) in the Institute for Computational and Mathematical Engineering at Stanford University, advised by Prof Emma Brunskill and Mykel J. Kochenderfer. His PhD dissertation investigated modern Reinforcement Learning challenges such as exploration, function approximation, adaptivity, and learning from offline data. His work was supported by a Total Innovation Fellowship and his PhD thesis was awarded the Gene Golub Outstanding Dissertation Award from his department. Andrea’s background is in mechanical engineering. Before Stanford, he worked as a software developer in high-performance computing, as well as at the von Karman Institute for Fluid Dynamics, a NATO-affiliated international research establishment.
Wednesday, April 12, 2023
Deconstructing Models and Methods in Deep Learning
Pavel Izmailov, final year PhD student in Computer Science at New York University, working with Andrew Gordon Wilson
Time: 1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: Machine learning models are ultimately used to make decisions in the real world, where mistakes can be incredibly costly. We still understand surprisingly little about neural networks and the procedures that we use to train them, and, as a result, our models are brittle, often rely on spurious features, and generalize poorly under minor distribution shifts. Moreover, these models are often unable to faithfully represent uncertainty in their predictions, further limiting their applicability. In this talk, Izmailov will present works on neural network loss surfaces, probabilistic deep learning, uncertainty estimation and robustness to distribution shifts. In each of these works, they aim to build foundational understanding of models, training procedures, and their limitations, and then use this understanding to develop practically impactful, interpretable, robust and broadly applicable methods and models.
Pavel Izmailov is primarily interested in understanding and improving deep neural networks. In particular, his interests include out of distribution generalization, probabilistic deep learning, representation learning and large models. Izmailov is also excited about generative models, uncertainty estimation, semi-supervised learning, language models and other topics. Recently, his work on Bayesian model selection was recognized with an outstanding paper award at ICML 2022.
Thursday, April 6, 2023
(Un)trustworthy Machine Learning: How to Balance Security, Accuracy, and Privacy
Eugene Bagdasaryan, CS PhD candidate at Cornell Tech and an Apple AI/ML PhD Scholar advised by Vitaly Shmatikov and Deborah Estrin
Time: 10:00 AM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: Machine learning methods have become a commodity in the toolkits of both researchers and practitioners. For performance and privacy reasons, new applications often rely on third-party code or pretrained models, train on crowd-sourced data, and sometimes move learning to users’ devices. This introduces vulnerabilities such as backdoors, (i.e., unrelated tasks that the model may unintentionally learn when an adversary controls parts of the training data or pipeline). In this talk, Bagdasaryan will identify new threats to ML models and propose approaches that balance security, accuracy, and privacy without disruptive changes to the existing training infrastructures.
Eugene Bagdasaryan is a doctoral candidate at Cornell University, where he is advised by Vitaly Shmatikov and Deborah Estrin. He studies how machine learning systems can fail or cause harm and how to make these systems better. His research has been published at security and privacy and machine learning venues and has been recognized by the Apple Scholars in AI/ML PhD fellowship.
Monday, April 3, 2023
Operationalizing Individual Fairness
Mikhail Yurochkin, Research Scientist at IBM Research and MIT-IBM Watson AI Lab.
1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: Societal applications of ML proved to be challenging due to algorithms replicating or even exacerbating biases in the training data. In response, there is a growing body of research on algorithmic fairness that attempts to address these issues, primarily via group definitions of fairness. In this talk, Yurochkin will illustrate several shortcomings of group fairness and present an algorithmic fairness pipeline based on individual fairness (IF). IF is often recognized as the more intuitive and desirable notion of fairness: we want ML models to treat similar individuals similarly. Despite the benefits, challenges in formalizing the notion of similarity and enforcing equitable treatment prevented the adoption of IF. He will present work addressing these barriers via algorithms for learning the similarity metric from data and methods for auditing and training fair models utilizing the intriguing connection between individual fairness and adversarial robustness. Finally, Yurochkin will discuss connections between algorithmic fairness and the emerging topic of robustness to distribution shifts and opportunities for future research in these areas.
Mikhail Yurochkin: Before joining IBM, he completed PhD in Statistics at the University of Michigan, advised by XuanLong Nguyen. Mikhail develops methods for reliable and inclusive adoption of ML and AI in practice using geometric ideas and tools from optimal transport, Bayesian modeling, and mathematical optimization. He led the development of the first open-source Python library for individual fairness (inFairness) and presented several tutorials on algorithmic fairness. He is also known for his work on model fusion and federated learning.
Thursday, March 30, 2023
Overcoming Obstacles for Reliable Reinforcement Learning in the Wild
Xuezhou Zhang, Associate Research Scholar at the ECE Department, Princeton University, working with Prof. Mengdi Wang.
Time: 11:00am | Location: CDS 1750 | 665 Commonwealth Ave, Boston
Abstract: Reinforcement Learning (RL) is a form of machine learning where an autonomous agent learns how to behave in an unknown environment through active interactions. Unlike traditional supervised learning, the agent has control over its environment and must gather feedback to make predictions. Despite its potential, RL has faced challenges in achieving success in real-life applications. In this talk, Zhang will describe three main obstacles for RL and offer solutions.
The first challenge is data efficiency. Existing deep RL algorithms are notoriously data-inefficient, requiring large amounts of interactions to find a good policy. Zhang will discuss how representation learning can provide a fundamental solution to this problem, both in theory and in practice. In particular, he will talk about how a low-dimensional representation can be learned from high-dimensional observations in an online fashion, and how such representations can facilitate efficient learning in current and future tasks.
The second challenge is security and robustness. To learn from raw interactions, an RL agent must be resilient against noisy data, distribution shifts, and potential adversarial attacks. He will talk about how wisdom from traditional robust statistics can help us transform deep RL algorithms to achieve robust learning even against the strongest adversaries.
The last challenge is to achieve efficient human-AI interaction. A machine agent deployed into the physical world will inevitably interact with other agents including both machine agents and humans. Zhang will present research findings on how learning from human feedback differs from learning from environmental feedback.
Xuezhou (Jack) Zhang is a postdoc associate in the ECE department and Center for Statistics and Machine Learning at Princeton University, working with Prof. Mengdi Wang. Before that, he received a Ph.D. in Computer Sciences from the University of Wisconsin-Madison, advised by Prof. Jerry Zhu. He was a visiting scholar in the Learning and Games program at Simons Institute for the Theory of Computing in Spring 2022. His research focuses on the theory and applications of interactive machine learning. More information can be found on his website.
Monday, March 27, 2023
Learning from the Interconnected World with Graphs
Jiaxuan You, CS Ph.D. from Stanford University, advised by Prof. Jure Leskovec
1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: The fact that our world is fundamentally interconnected presents unique challenges for modern data-driven research. In this talk, You will present his research on investigating the interconnected world through the lens of graphs. Specifically, he will present his research on accelerating graph AI research by systematically investigating the design space and task space for graph deep learning. The research has democratized graph AI for domain experts and helped them with scientific discoveries. Next, You will demonstrate my pioneering research in deep graph generative models which can generate novel realistic graph structures toward desirable objectives. This line of work has broad applications in molecule design and drug discovery. Lastly, he will show that graphs can further power AI problems in general domains. You will specifically cover my research in representing neural networks as relational graphs, which advances the design and understanding of deep neural networks and connects to network science and neuroscience. Overall, the talk will outline the promising path toward bridging interdisciplinary research and extending the frontiers of AI with graphs.
Jiaxuan You's research investigates scientific and industrial problems through the lens of graphs and develops graph AI methods to solve these problems. He has published 12 first-author papers in NeurIPS, ICML, ICLR, AAAI, KDD, and WWW, many of which are widely recognized. He has created or co-led multiple open-source software with over 20,000 combined GitHub stars. Jiaxuan has received multiple prestigious awards, including a JPMorgan Chase Ph.D. Fellowship, AAAI Best Student Paper Award, World Bank Best Big Data Solution, and Outstanding TA Award from Stanford CS. He was the lead organizer of NeurIPS New Frontiers in Graph Learning Workshop, a co-organizer of the Stanford Graph Learning Workshop, and a program committee member at numerous top-tier AI conferences and journals. His Ph.D. research further leads to a startup Kumo AI which demonstrates significant real-world impact.
Wednesday, March 22, 2023
Learning Systems in Adaptive Environments: Theory, Algorithms and Design
Aldo Pacchiano, Postdoctoral Researcher at Microsoft Research NYC
1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: Recent years have seen great successes in the development of learning algorithms in static predictive and generative tasks, where the objective is to learn a model that performs well on a single test deployment and in applications with abundant data. Comparatively less success has been achieved in designing algorithms for deployment in adaptive scenarios where the data distribution may be influenced by the choices of the algorithm itself, the algorithm needs to adaptively learn from human feedback, or the nature of the environment is rapidly changing. These are some of the most important challenges in the development of ML driven solutions for technologies such as interactive web systems, ML driven scientific experimentation, and robotics. To fully realize the potential of these technologies we will necessitate better ways of identifying problem domains and designing algorithms for adaptive learning.
To achieve this, in this talk, Pacchiano will propose adopting a systems view of adaptive learning mechanisms along with the following algorithm design considerations 1) development of sample efficient and tractable algorithms, 2) generalization to unseen domains via effective knowledge transfer and 3) human centric decision making. He will give an overview of his work along each of these axes and introduce a variety of open problems and research directions inspired by this conceptual framing.
Aldo Pacchiano obtained his PhD at UC Berkeley where he was advised by Peter Bartlett and Michael Jordan. His research lies in the areas of Reinforcement Learning, Online Learning, Bandits and Algorithmic Fairness. He is particularly interested in furthering our statistical understanding of learning phenomena in adaptive environments and use these theoretical insights and techniques to design efficient and safe algorithms for scientific, engineering, and large-scale societal applications.
Wednesday, March 22, 2023
How Many Deaths Has the COVID-19 Pandemic Really Caused in the U.S.? A Look at Current Evidence
Andrew Stokes, Boston University Assistant Professor (SPH) Global Health, (CAS) Sociology
4:00 PM | Location: CDS 1101 | 665 Commonwealth Ave, Boston
Refreshments and Conversation start at 3:30 PM
Abstract: Excess mortality has emerged as an important metric for monitoring all-cause and cause-specific mortality during the Covid-19 pandemic. Prior studies of excess mortality have primarily focused on national and regional trends but revealing the mortality impact of the Covid-19 pandemic at finer spatial-temporal resolutions is vital to developing targeted policy responses. In the present talk, Stokes discuss Bayesian small area estimation methods for calculating excess mortality for 3,128 counties in the United States over each month of the pandemic and highlight how spatial-temporal modeling of excess mortality can be used to quantify potential underreporting of Covid-19 deaths across local jurisdictions. He will also discuss time-series modeling approaches, including dynamic harmonic regression models with autoregressive integrated moving average (ARIMA), and highlight how these methods can be used to enrich excess mortality modeling and shed light on the drivers of increases in pandemic mortality from causes other than Covid-19. Finally, he will discuss how this modeling infrastructure may be used to create an early warning system for identifying anomalous mortality patterns associated with future pandemics and other public health emergencies.
Andrew Stokes: PhD is a demographer and sociologist with expertise in population health and aging. Through his research and dissemination efforts, he strives to reveal the social and structural factors that influence health across the life course, inform public health policies that center health equity, and contribute to evidence-based reforms of public health and health care systems. His research portfolio includes research on (1) spatial-temporal trends in excess mortality during the Covid-19 pandemic, (2) the determinants of long-term mortality and life expectancy trends, (3) chronic disease, pain, and disability across the life course, (4) substance use, tobacco, and e-cigarettes, and (5) access to care for chronic diseases in low- and middle-income countries. Founded in response to the Covid-19 pandemic in 2020, he created and leads a project focused on quantifying uncounted Covid-19 deaths in local communities across the United States. Through peer-reviewed research, media collaborations, and public data dashboards, the project seeks to reveal the hidden death toll of the pandemic and inform public health policies that center equity. Dr. Stokes received his B.A. in Environmental Studies from Bates College, M.A. in Demography from the University of Pennsylvania, and PhD in Demography and Sociology from the University of Pennsylvania. Prior to his doctoral studies, he was a post-bachelor fellow at the Harvard Initiative for Global Health in Cambridge, MA.
Yannis Paschalidis is a Distinguished Professor in the College of Engineering at Boston University with joint appointments in the Department of Electrical and Computer Engineering, the Division of Systems Engineering, and the Department of Biomedical Engineering. He is also a Founding Professor of Computing & Data Sciences. He is the Director of the Rafik B. Hariri Institute for Computing and Computational Science & Engineering – a Boston University federation of several centers and initiatives which acts as a catalyst and convergence accelerator for interdisciplinary research in this broad space.
Monday, March 20, 2023
Human-AI Interaction Under Societal Disagreement
Mitchell L. Gordon, Computer Science PhD student at Stanford University in the Human-Computer Interaction group, advised by Michael Bernstein and James Landay
1:30 PM | Location: CDS 1646 (in-person event) | 665 Commonwealth Ave, Boston
Abstract: Whose voices — whose labels — should a machine learning algorithm learn to emulate? For AI tasks ranging from online comment toxicity detection to poster design to medical treatment, different groups in society may have irreconcilable disagreements about what constitutes ground truth. Today’s supervised machine learning pipeline typically resolves these disagreements implicitly by majority vote over annotators’ opinions. This majoritarian procedure abstracts individual people out of the pipeline and collapses their labels into an aggregate pseudo-human, ignoring minority groups’ labels. In this talk, Gordon will present Jury Learning: an interactive AI architecture that enables developers to explicitly reason over whose voice a model ought to emulate through the metaphor of a jury. Through our exploratory interface, practitioners can declaratively define which people or groups, in what proportion, determine the classifier's prediction. To evaluate models under societal disagreement, he will also present The Disagreement Deconvolution: a metric transformation showing how, in abstracting away the individual people that models impact, current metrics dramatically overstate the performance of many user-facing tasks. These components become building blocks of a new pipeline for encoding our goals and values in human-AI systems, which strives to bridge the principles of HCI with the realities of machine learning.
Mitchell L. Gordon designs interactive systems and evaluation approaches that bridge the principles of human-computer interaction with the realities of machine learning. His work has won awards at top conferences in human-computer interaction and artificial intelligence, including a Best Paper award at CHI and an Oral at NeurIPS. He is supported by an Apple PhD Fellowship in AI/ML, and has interned at Apple, Google, and CMU HCII.
Monday, March 13, 2023
Behavioral Algorithms for Biological Learning and Decision-making
Gautam Reddy, PhD in Physics from the University of California San Diego, advised by Massimo Vergassola
1:30pm | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: Living systems sense their physical environment and process this information to interact back with the environment. This continual loop that iterates between sensing, computation and action drives complex behaviors. An enduring challenge is to identify the learning rules that enable animals to learn the structure of their environment and figure out algorithms to navigate them. In this talk, Dr. Reddy will highlight recent attempts to develop quantitative, phenomenological models of animal decision-making using high-throughput behavioral data and reinforcement learning. He will focus on rodent navigation in structured environments, to demonstrate how these methods can guide experimental design and how animal behavior can in turn motivate new machine learning paradigms.
Gautam Reddy: After two years as an independent NSF-Simons MathBio Fellow at Harvard, he is currently a Research Scientist at NTT Physics & Informatics Labs and the Center for Brain Science at Harvard. Gautam develops physics-inspired theory and methods for examining algorithmic aspects of biological computation. Specific contexts he has worked on include thermal soaring, olfactory navigation, microbial evolution and rodent learning in structured environments.
Wednesday, March 15, 2023
Designing and Analyzing Machine Learning Algorithms in the Presence of Strategic Behavior
Hanrui Zhang, PhD student at Carnegie Mellon University, advised by Vincent Conitzer
1:30pm | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: Machine learning algorithms now play a major part in all kinds of decision-making scenarios. When the stakes are high, self-interested agents --- about whom decisions are being made --- are increasingly tempted to manipulate the machine learning algorithm, in order to better fulfill their own goals, which are generally different from the decision maker's. This highlights the importance of making machine learning algorithms robust against manipulation. In this talk, Zhang will focus on generalization (i.e., the bridge between training and testing) in strategic classification: Traditional wisdom suggests that a classifier trained on historical observations (i.e., the training set) usually also works well on future data points to be classified (i.e., the test set). He will show how this very general principle fails when agents being classified strategically respond to the classifier, and present an intuitive fix that leads to provable (and in fact, optimal) generalization guarantees under strategic manipulation. Zhang will then discuss the role of incentive-compatibility in strategic classification, and present experimental results that illustrate how the theoretical results can guide practice. If time permits, he will discuss distinguishing strategic agents with samples, and/or dynamic decision making with strategic agents.
Hanrui Zhang's work won the Best Student Paper Award at the European Symposia on Algorithms (ESA), and a Honorable Mention for Best Paper Award at the AAAI Conference on Human Computation and Crowdsourcing (HCOMP). He received his bachelor's degree in Yao's Class, Tsinghua University, where he won the Outstanding Undergraduate Thesis Award.
Thursday, March 2, 2023
Towards Responsible Machine Learning in Societal Systems
Lydia Liu, PhD in Electrical Engineering and Computer Sciences from the University of California, Berkeley and postdoctoral researcher in Computer Science at Cornell University
11:00 AM | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: Machine learning systems are deployed in consequential domains such as education, employment, and credit, where decisions have profound effects on socioeconomic opportunity and life outcomes. High stakes decision settings present new statistical, algorithmic, and ethical challenges. In this talk, we examine the distributive impact of machine learning algorithms in societal contexts, and investigate the algorithmic and sociotechnical interventions that bring machine learning systems into alignment with societal values---equity and long-term welfare. First, we study the dynamic interactions between machine learning algorithms and populations, for the purpose of mitigating disparate impact in applications such as algorithmic lending and hiring. Next, we consider data-driven decision systems in competitive environments such as markets, and devise learning algorithms to ensure efficiency and allocative fairness. We end by outlining future directions for responsible machine learning in societal systems that bridge the gap between the optimization of predictive models and the evaluation of downstream decisions and impact.
Monday, February 27, 2023
Using Large Datasets and Machine Learning to Model Human Behavior
Joshua Peterson, PhD in Cognition, Brain & Behavior from the University of California, Berkeley and postdoctoral researcher in the Department of Computer Science at Princeton University
1:30 PM | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: Humans are complex and their specific behaviors can be resistant to scientific explanation despite decades of study. In this talk, I demonstrate cases where data-driven machine learning can be used to outpredict scientific models of even some of the most well-studied domains of human behavior. This success thus presents a new challenge: to understand the implicit theories about behavior that ML models are discovering. To this end, I present a novel method wherein carefully chosen inductive biases can be leveraged to produce interpretable scientific theories with no loss in prediction performance. When applied to the domain of decision making, this method results in a revision to a Nobel-winning theory in behavioral economics. These results lead me to a broader conclusion that modern tools from computer science have the potential to transform the science of the mind in the 21st century.
Wednesday, February 22, 2023
Socially Responsible & Factual Reasoning for Equitable AI Systems
Saadia Gabriel, PhD candidate, Paul G. Allen School of Computer Science & Engineering, University of Washington
1:30 PM | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: Understanding the implications underlying a text is critical to assessing its impact. This requires endowing artificial intelligence (AI) systems with pragmatic reasoning, for example to infer that the statement “Epidemics and cases of disease in the 21st century are “staged”” relates to unfounded conspiracy theories. This presentation will outline how shortcomings in the ability of current AI systems to reason about pragmatics leads to inequitable detection of false or harmful language, and demonstrate how these shortcomings can be addressed by imposing human-interpretable structure on deep learning architectures using insights from linguistics.
Thursday, February 16, 2023
Wei Jin, PhD candidate, Michigan State University
12:30 PM | Location: CDS 1646 | 665 Commonwealth Avenue, Boston
Abstract: This presentation will provide a fresh perspective on enhancing graph inputs, graph neural networks (GNNs) by optimizing the graph data, rather than designing new models. Specifically, Jin will present a model-agnostic framework which improves prediction performance by enhancing the quality of an imperfect input graph. Then show how to significantly reduce the size of a graph dataset while preserving sufficient information for GNN training.
Monday, January 30, 2023
Towards Secure and Regulated Machine Learning Systems
Emily Wenger, Ph.D. candidate, SAND Lab, University of Chicago
Abstract: This presentation highlighted two key areas of Wegner’s work: vulnerabilities in and caused by ML models and a novel attack discovered against computer vision models. Wegner explored building practical tools that protect models and empower users, highlighted a privacy tool she developed to disrupt unwanted facial recognition, followed by discussion of her vision for the future of secure and regulated ML.