CDS Spring 2025 Colloquium Series

The CDS Colloquium Series was developed to build an intellectual community within and beyond the academic unit. Since its inception, the series has welcomed dozens of scholars and is intended for CDS faculty, staff, and students. However, we welcome interest from across the Boston University campus and beyond.

To view all upcoming lectures, events, and programs, visit the CDS Calendar


Headshot of Valentina PyatkinTuesday, February 18, 2025

Training Precise Language Models for Imprecise Humans

Valentina Pyatkin, a postdoctoral researcher at the Allen Institute for AI and the University of Washington, advised by Prof. Yejin Choi.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: This talk examines methods for enhancing language model capabilities through post-training. While large language models have led to major breakthroughs in natural language processing, significant challenges persist due to the inherent ambiguity and underspecification in language. I will present a spectrum ranging from underspecification (preference modeling) to full specification (precise instruction following with verifiable constraints), and propose modeling approaches to increase language models’ contextual robustness and precision.

Valentina will demonstrate how models can become more precise instruction followers through synthetic data, preference tuning, and reinforcement learning from verifiable rewards. And she will address constrained instruction following generalization challenges and present post-training methods for improvement. On the preference data side, she will illustrate patterns of divergence in annotations, showing how disagreements stem from underspecification, and propose alternatives to the Bradley-Terry reward model for capturing pluralistic preferences.

The talk concludes by connecting underspecification and reinforcement learning through a novel method: reinforced clarification question generation, which helps models obtain missing contextual information that is consequential for making predictions. Throughout the presentation, Valentina will synthesize these research threads to demonstrate how post-training approaches can improve model steerability and contextual understanding when facing underspecification.

Bio: Valentina Pyatkin is a postdoctoral researcher at the Allen Institute for AI and the University of Washington, advised by Prof. Yejin Choi. She is additionally supported by an Eric and Wendy Schmidt Postdoctoral Award. She obtained her PhD in Computer Science from the NLP lab at Bar Ilan University. Her work has been awarded an ACL Outstanding Paper Award and the ACL Best Theme Paper Award. During her doctoral studies, she conducted research internships at Google and the Allen Institute for AI, where she received the AI2 Outstanding Intern of the Year Award. She holds an MSc from the University of Edinburgh and a BA from the University of Zurich. Valentina's research focuses on post-training and the adaptation of language models, specifically for making them better semantic and pragmatic reasoners. https://valentinapy.github.io/


Headshot of Sunnie KimThursday, February 20, 2025

Advancing Responsible AI with Human-Centered Evaluation

Sunnie S. Y. Kim, PhD candidate in Computer Science at Princeton University, advised by Olga Russakovsky.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: As AI technologies are increasingly transforming how we live, work, and communicate, AI evaluation must take a human-centered approach to realistically reflect real-world performance and impact. In this talk, Sunnie will discuss how to advance human-centered evaluation, and subsequently, responsible development of AI, by integrating knowledge and methods from AI and HCI. First, using explainable AI as an example, she will highlight the challenges and necessity of human (as opposed to automatic) evaluation. Second, she will illustrate the importance of contextualized evaluation with real users, revisiting key assumptions in explainable AI research. Finally, Sunnie will present empirical insights into human-AI interaction, demonstrating how users perceive and act upon common AI behaviors (e.g., LLMs providing explanations and sources). The talk will conclude by discussing the implications of these findings and future directions for responsible AI development.

Bio: Sunnie S. Y. Kim is a PhD candidate in Computer Science at Princeton University advised by Olga Russakovsky. She works on responsible AI and human-AI interaction — specifically, on improving the explainability and fairness of AI systems and helping people have appropriate understanding and trust in them. Her research has been published in both AI and HCI venues (e.g., CVPR, ECCV, CHI, FAccT), and she has organized multiple workshops connecting the two communities. She has been recognized by the NSF GRFP, Siebel Scholars, and Rising Stars in EECS, and has interned at Microsoft Research with the FATE group. Prior to graduate school, she received a BSc degree in Statistics and Data Science from Yale University and spent a year at Toyota Technological Institute at Chicago. https://sunniesuhyoung.github.io/


Headshot of Naomi SaphraMonday, February 24, 2025

Understanding Language Models by Understanding Training

Naomi Saphra, a Research Fellow at the Kempner Institute at Harvard University

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: LMs work better than anyone could have predicted just five years ago. But when do they work—and when don’t they? How do they work—and how do they fail? Why do they work—and why do they misbehave? This last question—why?—cannot be answered only by inspecting trained LMs. We must understand the underlying factors that produce LM behavior, an understanding grounded in the training process. For a given architecture, training is a recipe with three ingredients: time, data, and luck. Naomi will discuss these factors through controlled experiments inspecting and manipulating training. These experiments answer fundamental questions about why language models learn. How do training breakthroughs produce language competence? How can training data composition determine model capabilities? And when does output behavior depend on random initialization? By answering these questions, we can expose fundamental truths about why modern deep learning works so well, and even uncover the nature of reasoning itself.

Bio: Naomi Saphra is a Research Fellow at the Kempner Institute at Harvard University working to understand NLP training dynamics: how models learn to encode linguistic patterns or other structures, how generalization develops, and how we can introduce useful inductive biases into the training process. She has a particular interest in applying frameworks from evolutionary biology to understand neural networks. Recently, Dr. Saphra has become interested in fish. Previously, she earned a PhD from the University of Edinburgh on Training Dynamics of Neural Language Models and worked at NYU, Google, and Facebook. For fun, she writes historical and meta-scientific surveys of the state of machine learning. Outside of research, she plays roller derby under the name Gaussian Retribution, performs standup comedy, and shepherds disabled programmers into the world of code dictation.


Headshot of Luhuan WuWednesday, February 26, 2025

Probabilistic Inference for Generative Models: Enabling Scientific Discovery with Statistical Insights

Luhuan Wu, PhD candidate in the Department of Statistics at Columbia University, where she is co-advised by Profs David Blei and John Cunningham.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: Generative models extensively trained on domain data hold immense potential for scientific discovery, but unlocking their utility requires principled statistical tools to extract meaningful insights. In this talk, Luhan will present her work developing conditional inference methods for diffusion models — a class of generative models powering breakthroughs in protein design, image generation, and beyond. By leveraging the sequential Monte Carlo framework, her approach enables efficient and accurate sampling of outputs from pretrained diffusion models that satisfy domain-specific constraints. She will demonstrate its success in a protein design task, where the method generates protein structures with desired functional segments. Building on this work, she will outline her vision to establish a scalable and reliable probabilistic machine learning framework that bridges statistics, generative models, and modern scientific challenges.

Bio: Luhuan Wu is a PhD candidate in the Department of Statistics at Columbia University, where she is co-advised by Profs. David Blei and John Cunningham. Her research focuses on developing probabilistic machine learning methods to address challenges in modern scientific applications, including astrophysics, biology, and protein science. Her work spans large-scale spatio-temporal modeling, approximate inference, deep generative models, uncertainty quantification, and Bayesian modeling.


Past Talks

Headshot of Meena JagadeesanThursday, February 13, 2025

Steering Machine Learning Ecosystems of Interacting Agents

Meena Jagadeesan, a 5th year PhD student in Computer Science at UC Berkeley, where she is advised by Michael I. Jordan and Jacob Steinhardt.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: Modern machine learning models—such as LLMs and recommender systems—interact with humans, companies, and other models in a broader ecosystem. However, these multi-agent interactions often induce unintended ecosystem-level outcomes such as clickbait in classical content recommendation ecosystems, and more recently, safety violations and market concentration in nascent LLM ecosystems.

In this talk, Meena will discuss her research on characterizing and steering ecosystem-level outcomes. She will take an economic and statistical perspective on ML ecosystems, tracing outcomes back to the incentives of interacting agents and to the ML pipeline for training models. First, in LLM ecosystems, we show how analyzing a single model in isolation fails to capture ecosystem-level performance trends: for example, training a model with more resources can counterintuitively hurt ecosystem-level performance. To help steer ecosystem-level outcomes, we develop technical tools to assess how proposed policy interventions affect market entry, safety compliance, and user welfare. Then, turning to content recommendation ecosystems, we characterize a feedback loop between the recommender system and content creators, which shapes the diversity and quality of the content supply. Finally, she will present a broader vision of ML ecosystems where multi-agent interactions are steered towards the desired algorithmic, market, and societal outcomes.

Bio: Meena Jagadeesan is a 5th year PhD student in Computer Science at UC Berkeley, where she is advised by Michael I. Jordan and Jacob Steinhardt. Her research investigates multi-agent interactions in machine learning ecosystems from an economic and statistical perspective. She has received an Open Philanthropy AI Fellowship and a Paul and Daisy Soros Fellowship. https://mjagadeesan.github.io/


Monday, February 10, 2025

Modern Foundations of Social Prediction

Juan Perdomo, a postdoctoral fellow at the Harvard Center for Research on Computation and Society, advised by Cynthia Dwork.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: Machine learning excels at pattern recognition. Yet, when we deploy learning algorithms in social settings, we do not just aim to detect patterns; we use predictions to shape outcomes. This dynamic interplay, where we build systems using historical data to influence future behavior, underscores the role of prediction as both a lens and engine of social patterns. It also inspires us as researchers to explore new questions about which patterns we can influence, how to design prediction systems, and how to evaluate their impacts on society.

Juan will begin by presenting insights from his work on performative prediction: a learning-theoretic framework that formalizes the dynamic aspects of social prediction. In the second half, Juan will present an empirical case study evaluating the impact of a risk prediction tool used to allocate interventions to hundreds of thousands of public school students each year. He'll end with some discussion of future work and the challenges that lie ahead.

Bio: Juan earned his PhD from the University of California, Berkeley, where he was co-advised by Peter Bartlett and Moritz Hardt. His research focuses on the theoretical and empirical foundations of machine learning in society. Website: https://jcperdomo.org/


Mark Zhao BU CDS Colloquium Series

Tuesday, January 21, 2025

Thinking Outside the GPU: Systems for Scalable Machine Learning Pipelines

Mark Zhao is a final-year PhD candidate at Stanford University, advised by Christos Kozyrakis.

Time: 10:00 – 11:00 AM | Location: CDS 1646 (in-person event)

Abstract: Scalable and efficient machine learning (ML) systems have been instrumental in fueling recent advancements in ML capabilities. However, further scaling these systems requires more than simply increasing the number and performance of accelerators. This is because modern ML deployments rely on complex pipelines composed of many diverse and interconnected systems.

In this talk, Mark will emphasize the importance of building scalable systems across the entire ML pipeline. In particular, Mark will explore how large-scale ML training pipelines, including those deployed at Meta, require distributed data storage and ingestion systems to manage massive training datasets. Optimizing these data systems is essential as data demands continue to grow. To achieve this, Mark will demonstrate how synergistic optimizations across the training data pipeline can unlock performance and efficiency gains beyond what isolated system optimizations can achieve. While these synergistic optimizations are critical, deploying them requires navigating a large system design space. To address this challenge, Mark will introduce Cedar, a framework that automates the optimization and orchestration of ML data processing for diverse training workloads. Finally, he will discuss further opportunities in advancing the scalability, security, and capabilities of the hardware and software systems that continue to drive increasingly sophisticated ML training and inference pipelines.

Bio: Mark's research builds systems for end-to-end machine learning deployments by leveraging tools across the computing stack, including computer systems, computer architecture, security, databases, and machine learning. He has received an IEEE S&P Distinguished Practical Paper Award, a Top Pick in Hardware and Embedded Security Award, and an MLCommons ML and Systems Rising Star Award. His work is generously supported by a Stanford Graduate Fellowship and a Meta PhD. Fellowship in AI System SW/HW Co-Design. Website: https://web.stanford.edu/~myzhao/.


BU CDS Colloquium Xuhao ChenWednesday, January 22, 2025

Performance Engineering for Scalable AI and Big Data

Xuhao Chen is a Research Scientist at MIT CSAIL.

Time: 2:30 – 3:30 PM | Location: CDS 1646 (in-person event)

Abstract: AI applications are computationally expensive and hard to scale, which poses great challenges in computer system design.  In this talk, Xuhao will introduce his cross-stack performance engineering approach to address this challenge.  This approach involves performance optimization techniques and automation methods across different layers of the system stack, including algorithms, software, and hardware. He will share insights gained from his experiences building systems tailored for graph pattern mining (GPM), an important set of database and data mining algorithms. The first system is Scale-GPM, an algorithm, and software codesigned GPM system.

Next, he will discuss Pangolin, the first GPU-accelerated software programming system dedicated to GPM.  Complementing Pangolin is FlexMiner, a dedicated hardware accelerator engineered to further enhance the efficiency of GPM. Throughout the talk, he will showcase compelling results to underscore the effectiveness and the great potential of the cross-stack approach.

Bio: Dr. Chen is broadly interested in parallel systems and architectures, with a focus on AI and big-data applications. His recent work aims to make AI scalable by designing efficient algorithms, software systems, and hardware accelerators. His work has been published in VLDB, OSDI, ISCA, MICRO, ICS, etc. Website: https://www.csail.mit.edu/person/xuhao-chen.


Thursday, January 23, 2025

BU CDS Colloquium Kinan Dak AlbabPractical Privacy Compliance via New Systems and Abstractions

Kinan Dak Albab is a PhD candidate at Brown University advised by Malte Schwarzkopf.

Time: 3:30 – 4:30 PM | Location: CDS 1646 (in-person event)

Abstract: Data privacy has become a focal point for public discourse. In response, Data protection and privacy regulations have been enacted globally, including the GDPR and CCPA, and companies make various promises to end-users in their privacy policies. However, high-profile privacy violations remain commonplace, in part because complying with privacy regulations and policies is challenging for applications and developers.

This talk demonstrates Kinan and researchers can help developers achieve privacy compliance by designing new privacy-conscious systems and abstractions. This talk focuses on his work on Sesame (SOSP24), my system for end-to-end compliance with privacy policies in web applications. To provide practical guarantees, Sesame combines new static analysis for data leakage with advances in memory-safe languages and lightweight sandboxing, as well as standard industry practices like code review. Kinan's work in this area also includes K9db (OSDI23), a privacy-compliant database that supports compliance by construction with GDPR-style subject access requests. By creating privacy abstractions at the systems level, we can offer applications privacy guarantees by design, to simplify compliance and improve end-user privacy.

Bio: Kinan is interested in building real systems and practical tools to improve privacy in the real world using techniques from computer systems, cryptography, and programming languages. His software has been used in the real world to perform privacy-preserving analytics for the social good, and validate the next generation of SDN network switches at Google. Website:  https://www.babman.io/.


Monday, January 27, 2025

Headshot of Wenqi Jiang

Beyond Model Acceleration in Next-Generation Machine Learning Systems

Wenqi Jiang is a fifth-year PhD student at ETH Zurich, advised by Gustavo Alonso and Torsten Hoefler.

Time: 10:00 – 11:00 AM | Location: CDS 1646

Abstract: Despite the recent popularity of large language models (LLMs), the transformer neural network invented eight years ago has remained largely unchanged. It prompts the question of whether machine learning (ML) systems research is solely about improving hardware and software for tensor operations. In this talk, Wenqi will argue that the future of machine learning systems extends far beyond model acceleration. Using the increasingly popular retrieval-augmented generation (RAG) paradigm as an example, they will show that the growing complexity of ML systems demands a deeply collaborative effort spanning data management, systems, computer architecture, and ML.

Wenqi will present RAGO and Chameleon, two pioneering works in this field. RAGO is the first systematic performance study of retrieval-augmented generation. It uncovers the intricate interactions between vector data systems and models, revealing drastically different performance characteristics across various RAG workloads. To navigate this complex landscape, RAGO introduces a system optimization framework to explore optimal system configurations for arbitrary RAG algorithms. Building on these insights, Wenqi will introduce Chameleon, the first heterogeneous accelerator system for RAG. Chameleon combines LLM and retrieval accelerators within a disaggregated architecture. The heterogeneity ensures efficient serving of both LLM inference and retrievals, while the disaggregation enables independent scaling of different system components to accommodate diverse RAG workload requirements. Wenqi will conclude the talk by emphasizing the necessity of cross-stack co-design for future ML systems and the abundant of opportunities ahead of us.

Bio: Wenqi aims to enable more efficient, next-generation machine learning systems. Rather than focusing on a single layer in the computing stack, Wenqi's research spans the intersections of data management, computer systems, and computer architecture. His work has driven advancements in several areas, including retrieval-augmented generation (RAG), vector search, and recommender systems. These contributions have earned him recognition as one of the ML and Systems Rising Stars, as well as the AMD HACC Outstanding Researcher Award. Website: https://wenqijiang.github.io/.


TitleTuesday, January 28, 2025

Advancing Efficient and Trustworthy AI on the Edge

Jingwei Sun is a final-year PhD student in ECE at Duke University, advised by Prof. Yiran Chen.

Time: 10:00 – 11:00 AM | Location: CDS 1646

Abstract: Edge AI brings intelligence closer to users, enabling real-time, personalized interactions while maintaining data privacy. However, the increasing reliance on edge devices presents two significant challenges: ensuring efficient AI operations within resource-constrained environments and addressing trustworthiness concerns to safeguard sensitive on-device data. In this talk, Jingwei will first talk about efficient model training and personalization on edge devices, focusing on methods that eliminate the need for backpropagation to reduce computational and memory costs. Then, they will present works on enhancing privacy and robustness for Edge AI, introducing strategies to protect data integrity and defend against adversarial threats. These efforts address critical limitations and pave the way for efficient and trustworthy Edge AI solutions.

Bio: Jingwei's research focuses on efficient and trustworthy edge intelligent systems. His work appears in AI conferences such as NeurIPS, ICML, CVPR, and ICCV , as well as system conferences such as MobiCom, SenSys, and MLsys. He has received the Best Paper Award from the AAAI Spring Series Symposium 2024. Website: https://jingwei-sun.com/.

View Past Events