April 2025 - talks - lists.cs.princeton.edu

CS Colloquium speakers: week of February 26
by Emily C. Lawrence 12 May '25

12 May '25

CS Colloquium speakers Speaker: Jialin Ding, Amazon Web Services Date: Monday, February 26 Time: 12:30pm EST Location: CS 105 Host: Wyatt Lloyd Event page: https://www.cs.princeton.edu/events/26578 Register for live-stream online here: https://princeton.zoom.us/webinar/register/WN_10z_l8MYRr21ZOuPqA8Odw Title: Instance-Optimization: Rethinking Database Design for the Next 1000X Abstract: Modern database systems aim to support a large class of different use cases while simultaneously achieving high performance. However, as a result of their generality, databases often achieve adequate performance for the average use case but do not achieve the best performance for any individual use case. In this talk, I will describe my work on designing databases that use machine learning and optimization techniques to automatically achieve performance much closer to the optimal for each individual use case. In particular, I will present my work on instance-optimized database storage layouts, in which the co-design of data structures and optimization policies improves query performance in analytic databases by orders of magnitude. I will highlight how these instance-optimized data layouts address various challenges posed by real-world database workloads and how I implemented and deployed them in production within Amazon Redshift, a widely-used commercial database system. Bio: Jialin Ding is an Applied Scientist at AWS. Before that, he received his PhD in computer science from MIT, advised by Tim Kraska. He works broadly on applying machine learning and optimization techniques to improve data management systems, with a focus on building databases that automatically self-optimize to achieve high performance for any specific application. His work has appeared in top conferences such as SIGMOD, VLDB, and CIDR, and has been recognized by a Meta Research PhD Fellowship. To learn more about Jialin’s work, please visit [ https://jialinding.github.io/ | https://jialinding.github.io/ ] . CSML/CS/PSY Colloquium Speaker: Brenden Lake, New York University Date: Tuesday, February 27 Time: 12:30pm EST Location: CS 105 Host: Tom Griffiths Event page: https://csml.princeton.edu/events/csmlpsycs-seminar Register for live-stream online here: https://princeton.zoom.us/webinar/register/WN_aUfLfVQFSIqV2RDSYu4big Title: Towards more human-like learning in machines: Bridging the data and generalization gaps Abstract: There is an enormous data gap between how AI systems and children learn: The best LLMs now learn language from text with a word count in the trillions, whereas it would take a child roughly 100K years to reach those numbers through speech (Frank, 2023, "Bridging the data gap"). There is also a clear generalization gap: whereas machines struggle with systematic generalization, children can excel. For instance, once a child learns how to "skip," they immediately know how to "skip twice" or "skip around the room with their hands up" due to their compositional skills. In this talk, I'll describe two case studies in addressing these gaps. 1) The data gap: We train deep neural networks from scratch (using DINO, CLIP, etc.), not on large-scale data from the web, but through the eyes and ears of a single child. Using head-mounted video recordings from a child as training data (<200 hours of video slices over 26 months), we show how deep neural networks can perform challenging visual tasks, acquire many word-referent mappings, generalize to novel visual referents, and achieve multi-modal alignment. Our results demonstrate how today's AI models are capable of learning key aspects of children's early knowledge from realistic input. 2) The generalization gap: Can neural networks capture human-like systematic generalization? We address a 35-year-old debate catalyzed by Fodor and Pylyshyn's classic article, which argued that standard neural networks are not viable models of the mind because they lack systematic compositionality -- the algebraic ability to understand and produce novel combinations from known components. We'll show how neural networks can achieve human-like systematic generalization when trained through meta-learning for compositionality (MLC), a new method for optimizing the compositional skills of neural networks through practice. With MLC, neural networks can match human performance and solve several machine learning benchmarks. Given these findings, we'll discuss the paths forward for building machines that learn, generalize, and interact in more human-like ways based on more natural input. Related articles: Vong, W. K., Wang, W., Orhan, A. E., and Lake, B. M (2024). Grounded language acquisition through the eyes and ears of a single child. Science, 383, 504-511. Orhan, A. E., and Lake, B. M. (in press). Learning high-level visual representations from a child’s perspective without strong inductive biases. Nature Machine Intelligence. Lake, B. M. and Baroni, M. (2023). Human-like systematic generalization through a meta-learning neural network. Nature, 623, 115-121. Bio: Brenden M. Lake is an Assistant Professor of Data Science and Psychology at New York University. He received his M.S. and B.S. in Symbolic Systems from Stanford University in 2009, and his Ph.D. in Cognitive Science from MIT in 2014. He was a postdoctoral Data Science Fellow at NYU from 2014-2017. Brenden is a recipient of the Robert J. Glushko Prize for Outstanding Doctoral Dissertation in Cognitive Science and the MIT Technology Review 35 Innovators Under 35. His research was also selected by Scientific American as one of the 10 most important advances of 2016. Brenden's research focuses on computational problems that are easier for people than they are for machines, such as learning new concepts from just a few examples, learning by asking questions, learning by generating new goals, and learning by producing novel combinations of known components. CS Colloquium Speaker: Eric Mitchell, Stanford University Date: Thursday, February 29 Time: 12:30pm EST Location: CS 105 Host: Karthik Narasimhan Event page: https://www.cs.princeton.edu/events/26587 Register for live-stream online here: TBA Talk info TBA

1 24

Adithya Bhaskar will present his General Exam "Effective mechanistic interpretability via pruning" on Monday, May 5, 2025 at 3:00 PM in FC 008 and via zoom.
by CS Grad Department 30 Apr '25

30 Apr '25

Adithya Bhaskar will present his General Exam "Effective mechanistic interpretability via pruning" on Monday, May 5, 2025 at 3:00 PM in FC 008 and via zoom. Zoom link: https://princeton.zoom.us/j/98937358810 Committee Members: Danqi Chen (advisor), Sanjeev Arora, Benjamin Eysenbach Abstract: Mechanistic interpretability strives to explain language models bottom-up by understanding their components. In this talk, I will describe how pruning—removing less important parameters or connections in a model—can enable a new approach to interpretability. I will introduce a curious phenomenon as a case study: fine-tuning a language model with different random seeds can lead to similar in-domain performance but exhibit distinct patterns of generalization. Pruning the model reveals that even a single model contains many subnetworks that match its in-domain performance, but generalize differently. By connecting this result to a “grokking”-like phenomenon observed for out-of-domain generalization, I will develop a theory of why these subnetworks exist in the model, and what role they fulfill. Zooming out of this example, we will apply pruning to develop an efficient, accurate, and scalable tool to automatically discover transformer circuits—basic units of mechanistic interpretability. We develop Edge Pruning, a method that focuses on pruning edges rather than the parameters of a language model. Edge Pruning discovers circuits substantially more faithful to the model as measured by KL divergence while being more efficient and scaling to 100x larger models than prior approaches. We will see via an example how this tool enables future work to understand phenomena that only emerge in large (1B+ parameters) models, such as in-context learning. Our results imply that pruning holds great promise for interpretability—both as a paradigm and as a standalone method. Reading List: https://docs.google.com/document/d/1VJ5busQNlb-er_yt7-ae34x2UC0KXZh2irvR0vi… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Stamatis Alexandropoulos will present his General Exam "Princeton365: A Diverse Dataset with Accurate Camera Pose" on Monday, May 5, 2025 at 2:00 PM in FC 005 and via zoom.
by CS Grad Department 30 Apr '25

30 Apr '25

Stamatis Alexandropoulos will present his General Exam " Princeton365: A Diverse Dataset with Accurate Camera Pose " on Monday, May 5, 2025 at 2:00 PM in FC 005 and via zoom. Zoom link: https://princeton.zoom.us/my/stamalex Committee Members: Jia Deng (advisor), Felix Heide , Adam Finkelstein Abstract: Accurately evaluating SLAM systems often comes at the cost of limited data diversity, while diverse datasets usually lack precise camera pose. In this talk, I will introduce Princeton365, a new dataset designed to bridge this gap. It includes 365 videos captured using a 360° camera and calibration boards to provide accurate ground truth camera poses. The dataset features a wide range of scenarios—including indoor, outdoor, and object scanning scenes—and offers synchronized monocular and stereo RGB video, along with IMU data. I’ll also present a new scale-aware evaluation metric for SLAM based on optical flow induced by pose estimation errors, which enables more consistent comparison across different scenes than traditional metrics like Average Trajectory Error. Finally, I’ll introduce a novel benchmark for view synthesis that tackles challenging cases not handled by current datasets, such as non-Lambertian surfaces and full 360° camera trajectories. Reading List: https://docs.google.com/document/d/1iqBGxihNyFdlsjk5eCaNgA8EV0hOObt1Z_CN9JG… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Erich Liang will present his General Exam "ChangeIn: A Benchmark for Self-Calibration of Dynamic Intrinsics of Video Cameras" on Monday, May 5, 2025 at 10:30 AM in FC 202 and via zoom.
by CS Grad Department 30 Apr '25

30 Apr '25

Erich Liang will present his General Exam "ChangeIn: A Benchmark for Self-Calibration of Dynamic Intrinsics of Video Cameras" on Monday, May 5, 2025 at 10:30 AM in FC 202 and via zoom. Zoom link: https://princeton.zoom.us/j/8238970644 Committee Members: Jia Deng (advisor), Szymon Rusinkiewicz , Ellen Zhong Abstract: Accurately tracking camera intrinsics is crucial for achieving 3D understanding from 2D video. However, most 3D algorithms assume that camera intrinsics stay constant throughout a video, which is often not true for many real-world in-the-wild videos. A major obstacle in this field is a lack of dynamic camera intrinsics benchmarks--existing benchmarks typically offer limited diversity in scene content and intrinsics variation, and none provide per-frame intrinsic changes for consecutive video frames. In this paper, we present ChangeIn, a real-world benchmark that provides per-frame ground truth intrinsics annotations for videos with dynamic intrinsics. Compared to prior benchmarks, ChangeIn captures a wider range of intrinsic variations and scene diversity, featuring 143K annotated frames from 386 high-resolution indoor and outdoor videos with dynamic camera intrinsics. To ensure accurate per-frame intrinsics, we build a comprehensive look-up table of calibration experiments and extend the Kalibr toolbox to improve its accuracy and robustness. Using our benchmark, we evaluate existing baseline methods for predicting camera intrinsics and find that most struggle to achieve accurate predictions on videos with dynamic intrinsics. Reading List: https://docs.google.com/document/d/1pu3P9hWcthJRMvBSssPqWLYiq3QZCVfa4p_v50J… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Kathryn Wantlin will present her General Exam "Self-Supervised Multi-Task Exploration and RL" on Friday, May 2, 2025 at 3:00 PM in CS 401.
by CS Grad Department 29 Apr '25

29 Apr '25

Kathryn Wantlin will present her General Exam "Self-Supervised Multi-Task Exploration and RL" on Friday, May 2, 2025 at 3:00 PM in CS 401. Committee Members: Ryan Adams (advisor), Ben Eysenbach, Olga Russakovsky Abstract: Learning from demonstrations to imitate a single behavior in a given environment can require hundreds or thousands of expert (human) demonstrations and learning a distribution of tasks in an environment is even more prohibitively inefficient. This work presents a framework for achieving zero-shot generalization capabilities in a given environment without access to expert data. We replace the imitation learning training phase with a self-supervised exploration phase, training a goal-conditioned critic, policy and variational posterior for inferring goals, and at test time, are able to zero-shot infer a goal from a demonstration and then unroll the associated goal-conditioned policy to imitate the demonstration’s behavior with low regret. We make the key assumption that many behaviors are likely goal-reaching, an inductive bias that inferring an agent’s goal provides a good bias-variance tradeoff compared to inferring their reward function. During the self-supervised exploration phase, I learn temporally contrastive features and utilize several methods for automatic goal selection, comparing these results against oracle goal sampling. Our pipeline is compared against related methods for zero-shot imitation, showing competitive performance against goal-conditioned behavioral cloning across our benchmarked environments. Reading List: https://docs.google.com/document/d/1zhhjPl5Ei1f_awCVUBwduVqi6F1AHZ6iMiJdsiT… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Han Xu will present his General Exam " Relational Netkat" on May 2, 2025 at 10am in CS 402 and Zoom.
by Gradinfo 29 Apr '25

29 Apr '25

Han Xu will present his General Exam " Relational Netkat" on May 2, 2025 at 10am in CS 402 and Zoom. Committee: David Walker (adviser), Zachary Kincaid, Maria Apostolaki Zoom link: [ https://princeton.zoom.us/j/8011315368 | https://princeton.zoom.us/j/8011315368 ] Abstract: Modern networks are complex, dynamic, and large-scale, making network verification a critical challenge. Relational NetKAT extends traditional NetKAT by providing an expressive and efficient framework for reasoning about network behavior under path changes, forwarding table modifications, and network topology updates. It enables all-path verification, ensuring that network-wide invariants hold across diverse changes such as load balancing and routing updates. Compared to sample-based verification of Rela, which fails to cover all possible paths, Relational NetKAT provides a mathematically sound and scalable approach to checking network correctness. Reading List: Textbook: * Types and programming languages. MIT press, 2002. * Handbook of Model Checking Ch 7.8. Paper: Netkat, Kleene algebra and automata theory * Xieyang Xu, Yifei Yuan, Zachary Kincaid, Arvind Krishnamurthy, Ratul Mahajan, David Walker, and Ennan Zhai. 2024. Relational Network Verification. * Carolyn Jane Anderson, Nate Foster, Arjun Guha, Jean-Baptiste Jeannin, Dexter Kozen, Cole Schlesinger, and David Walker. 2014. NetKAT: semantic foundations for networks. * Nate Foster, Dexter Kozen, Mae Milano, Alexandra Silva, and Laure Thompson. 2015. A Coalgebraic Decision Procedure for NetKAT. * Damien Pous. 2015. Symbolic Algorithms for Language Equivalence and Kleene Algebra with Tests. BDD and Model Checking * Bryant Randy, 1992. Symbolic Boolean manipulation with ordered binary-decision diagrams * Kathi Fisler and Moshe Y. Vardi. 2002. Bisimulation Minimization and Symbolic Model Checking. Network Verification * H. Yang and S. S. Lam, "Real-time verification of network properties using Atomic Predicates," 2013 * Peyman Kazemian, George Varghese, and Nick McKeown. 2012. Header space analysis: static checking for networks. * Ahmed Khurshid, Wenxuan Zhou, Matthew Caesar, and P. Brighten Godfrey. 2012. Veriflow: verifying network-wide invariants in real time.

1 0

Minkyu Jeon will present his General Exam "CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM" on Friday, May 2, 2025 at 12:00 PM in Friend 009.
by CS Grad Department 29 Apr '25

29 Apr '25

Minkyu Jeon will present his General Exam " CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM " on Friday, May 2, 2025 at 12:00 PM in Friend 009 . Committee Members: Ellen D. Zhong (advisor), Felix Heide , Jia Deng Abstract: Cryo-electron microscopy (cryo-EM) is a powerful technique for determining high-resolution 3D biomolecular structures from imaging data. Its unique ability to capture structural variability has spurred the development of heterogeneous recon- struction algorithms that can infer distributions of 3D structures from noisy, unlabeled imaging data. Despite the growing number of advanced methods, progress in the field is hindered by the lack of standardized benchmarks with ground truth information and reliable validation metrics. Here, we introduce CryoBench, a suite of datasets, metrics, and benchmarks for heterogeneous reconstruction in cryo-EM. CryoBench includes five datasets representing different sources of heterogeneity and degrees of difficulty. These include conformational heterogeneity generated from designed motions of antibody complexes or sampled from a molecular dynamics simulation, as well as compositional heterogeneity from mixtures of ribosome assembly states or 100 common complexes present in cells. We then analyze state-of-the-art heterogeneous reconstruction tools, including neural and non-neural methods, assess their sensitivity to noise, and propose new metrics for quantitative evaluation. We hope that CryoBench will be a foundational resource for accelerating algorithmic development and evaluation in the cryo-EM and machine learning communities. Reading List: https://docs.google.com/document/d/1QqGrfGd_nDdNno5_lZNBbGTktW9NFxE_52wzXC4… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

You're Invited! AI Lab Distinguished Lecture Series: May 2
by Emily C. Lawrence 29 Apr '25

29 Apr '25

[ https://pli.princeton.edu/events/2025/ai-lab-distinguished-lecture-series-a… | AI Lab Distinguished Lecture Series ] “AI as a New Paradigm for Scientific Discovery”: Insights from Google DeepMind Pushmeet Kohli, VP of Research, Google DeepMind Friday May 2 at 3:30 p.m. in Computer Science 104 Pushmeet Kohli is a researcher who holds the position of vice president of research for Google DeepMind. Before joining DeepMind, Pushmeet was the director of research at the cognition group at Microsoft. During his 10 years at Microsoft, Pushmeet worked in Microsoft labs in Seattle, Cambridge and Bangalore and took a number of roles and duties, including being technical advisor to Rick Rashid, the Chief Research Officer of Microsoft. Pushmeet’s research revolves around intelligent systems and computational sciences, and he publishes in the fields of machine learning, computer vision, information retrieval, and game theory. His current research interests include 3D reconstruction and rendering, probabilistic programming, interpretable and verifiable knowledge representations from deep models. He is also interested in conversation agents for task completion, machine learning systems for healthcare and 3D rendering and interaction for augmented and virtual reality. Reception to follow

1 0

Ethan Tseng will present his FPO "Neural Proxy Functions for Computational Cameras and Displays" on Friday, May 2, 2025 at 2:30 PM in CS 105 and Zoom.
by CS Grad Department 28 Apr '25

28 Apr '25

Ethan Tseng will present his FPO "Neural Proxy Functions for Computational Cameras and Displays" on Friday, May 2, 2025 at 2:30 PM in CS 105 and Zoom. Location: Zoom link: https://princeton.zoom.us/j/97976329281 The members of Ethan’s committee are as follows: Examiners: Felix Heide (Adviser), Adam Finkelstein, Jia Deng Readers: Szymon Rusinkiewicz, Tian-Ming Fu A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Cameras and displays serve as portals between the physical world and the digital realm. We use cameras to take pictures of our environment and we use computer vision algorithms to analyze these digital images. These images and their associated analyses can then be shared with others through programmable displays ranging from smartphone screens to mixed reality headsets. However, despite their importance as gateways to the physical world, these hardware devices are still usually developed in isolation from the downstream software. We would like to be able to “learn” new cameras and displays in a manner similar to how modern machine learning models learn new functions by being trained on large volumes of data. Today’s machine learning models are typically trained using some form of gradient descent and the parameters of the model are progressively updated over many iterations. Extending the same design strategy to cameras and displays is challenging because gradients may not be available (e.g., proprietary black-box systems) or the gradients may be too computationally expensive to acquire (e.g., full wave physics models). This thesis introduces neural proxy functions to address and overcome these obstacles. Neural proxy functions act as differentiable approximations for systems that do not readily admit gradients. The gradients provided by neural proxy functions allow for the co-evolution of hardware and software, extending the reach of machine learning beyond the digital world. We demonstrate the utility of neural proxy functions through three research contributions: (1) Hyperparameter optimization of image signal processors that connect high-level visual machine learning (e.g., object detection) to low-level irradiance measurements recorded by a camera sensor. (2) Fast differentiable forward models for optical stacks consisting of both traditional optics and emerging nanophotonic elements. (3) Design of neural algorithms and optics for high-performance holographic displays.

1 0

Seth Karten will present his General Exam "Multi-Agent Frontiers with Foundation Models: Cooperative and Competitive Games" on May 6th, 2025 in Friend 125 at 12:30pm.
by Gradinfo 28 Apr '25

28 Apr '25

Seth Karten will present his General Exam "Multi-Agent Frontiers with Foundation Models: Cooperative and Competitive Games" on May 6th, 2025 in Friend 125 at 12:30pm. Committee: Chi Jin (adviser), Benjamin Eysenbach, Danqi Chen Title: "Multi-Agent Frontiers with Foundation Models: Cooperative and Competitive Games" Abstract: Recent advances in foundation models, particularly large language models (LLMs), have created unprecedented opportunities to enhance multi-agent reinforcement learning (MARL) systems by leveraging their rich world knowledge and reasoning capabilities. This examination explores two complementary approaches to integrating LLMs with MARL across different game-theoretic settings. In zero-sum environments, we introduce PokéChamp, an expert-level minimax language agent for Pokémon battles that replaces three key components of traditional search algorithms with LLM-based modules: player action sampling, opponent modeling, and value function estimation. PokéChamp achieves human-expert performance without additional training, demonstrating the potential of LLMs to effectively constrain search spaces using informative priors in partially observable environments. For general-sum scenarios, we present the LLM Economist, a novel framework that employs LLMs to create realistic economic simulations with varied agent populations. By formulating the problem as a Stackelberg game, we show how LLMs can model complex economic behaviors through in-context optimization of utility functions, generation of synthetic human preferences, and mechanism design for social welfare maximization. These approaches collectively demonstrate a powerful synergy: MARL architectures benefit from the rich knowledge and reasoning capabilities embedded in foundation models, while reinforcement learning planning techniques provide the grounding and optimization mechanisms necessary to overcome the limitations of foundation models in strategic decision-making contexts.

1 0