- talks - lists.cs.princeton.edu

Robotics Seminar TOMORROW at 11am
by Emily C. Lawrence 30 Apr '26

30 Apr '26

Hi everyone, We're excited to be hosting Professor Tom Dietterich from OSU this Friday, May 1 . The seminar will be held in CS 105 from 11 am - 12 pm . The seminar will be followed by lunch in EQuad F-wing. Title: Can We Make Machine Learning Safe for Safety-Critical Systems? Abstract : The impressive new capabilities of systems created using deep learning are encouraging engineers to apply these techniques in safety-critical applications such as medicine, aeronautics, and self-driving cars. This talk will discuss the ways that machine learning methodologies are changing to operate in safety-critical systems. These changes include (a) building high-fidelity simulators for the domain, (b) adversarial collection of training data to ensure coverage of the so-called Operational Design Domain (ODD) and, specifically, the hazardous regions within the ODD, (c) methods for verifying that the fitted models generalize well, and (d) methods for estimating the probability of harms in normal operation. There are many research challenges to achieving these. But we must do more, because traditional safety engineering only addresses the known hazards. We must design our systems to detect novel hazards as well. We adopt Leveson’s view of safety as an ongoing hierarchical control problem in which controls are put in place to stabilize the system against disturbances. Disturbances include novel hazards but also management changes such as budget cuts, staff turnover, novel regulations, and so on. Traditionally, it has been the human operators and managers who have provided these stabilizing controls. Are there ways that AI methods, such as novelty detection, near-miss detection, diagnosis and repair, can be applied to help the human organization manage these disturbances and maintain system safety? Bio : Dr. Dietterich is University Distinguished Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University. Dietterich is one of the pioneers of the field of Machine Learning and has authored more than 220 refereed publications and two books. His current research topics include robust artificial intelligence, robust human-AI systems, and applications in sustainability. Dietterich is the 2025 recipient of the Feigenbaum prize for applied AI and the 2024 recipient of the IJCAI Award for Research Excellence. Dietterich is also the recipient of the 2022 AAAI Distinguished Service Award and the 2020 ACML Distinguished Contribution Award, both recognizing his many years of service to the research community. He is a former President of the Association for the Advancement of Artificial Intelligence and the founding president of the International Machine Learning Society. Other major roles include Executive Editor of the journal Machine Learning, co-founder of the Journal for Machine Learning Research, and program chair of AAAI 1990 and NIPS 2000. He currently chairs the Computer Science Section of [ http://arxiv.org/ | arXiv.org ] .

1 0

Xinhao Liu will present his FPO "Computational Tools for Alignment and Integration of Spatial Multimodality Data" on Wednesday, May 6, 2026 at 1:30p in COS 402.
by CS Grad Department 29 Apr '26

29 Apr '26

Xinhao Liu will present his FPO "Computational Tools for Alignment and Integration of Spatial Multimodality Data" on Wednesday, May 6, 2026 at 1:30p in COS 402. The members of Xinhao’s committee are as follows: Examiners: Ben Raphael (Adviser), Yuri Pritykin, Ellen Zhong Readers: Olga Troyanskaya, Li Ding (Washington University in St. Louis) 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 the talk. Abstract follows below: Recent advances in spatial sequencing technologies, such as spatial transcriptomics, proteomics, and epigenomics, enable the in situ measurement of molecular features across intact tissue sections, providing new insights into the spatial organization of biological processes. As datasets grow in complexity, spanning multiple tissue slices, developmental timepoints, and molecular modalities, there is an increasing need for computational methods that can align and integrate spatial omics data into coherent, high-resolution representations of tissues and organs. Accurate alignment across slices is essential not only for reconstructing tissue architecture but also for comparing molecular programs across spatial scales, developmental stages, and experimental modalities. This thesis presents four computational methods for aligning and integrating spatial omics data across developmental timepoints and experimental modalities. First, we introduce PASTE2, a method for alignment and three-dimensional (3D) reconstruction of multislice spatial transcriptomics data that allows slices to overlap only partially. Next, we present DeST-OT, which aligns spatial transcriptomics slices collected at different developmental timepoints to capture tissue dynamics over time. We then introduce MOSAICField, a framework for aligning spatial slices across arbitrary combinations of experimental modalities, such as spatial transcriptomics, spatial proteomics, and histology images. Finally, we introduce OTVI, a machine learning framework that integrates multiple spatial modalities from related tissue slices into a common coordinate framework while learning biologically meaningful representations for each spatial location.

1 0

Zijian Qin will present his General Exam "Avicenna: Fail-Slow Fault-Tolerant Replicated State Machines Using Counterfactual Evaluation" on Weds 5/6/2026 at 9:30am at 194 Nassau St, R. 213.
by Gradinfo 29 Apr '26

29 Apr '26

Zijian Qin will present his General Exam "Avicenna: Fail-Slow Fault-Tolerant Replicated State Machines Using Counterfactual Evaluation" on Weds 5/6/2026 at 9:30am at 194 Nassau St, R. 213. Committee: Wyatt Lloyd (adviser), Jialin Ding, Mae Milano Abstract: Geo-distributed replicated state machines (RSMs) are at the heart of many production distributed systems, offering linearizability and fault tolerance via consensus protocols. Most existing protocols target crash fault tolerance, however, and are vulnerable to fail-slow faults, where a single slow replica can significantly degrade system latency. Existing protocols that tolerate fail-slow faults do so with much higher normal-case latency in geo-distributed settings. Avicenna is the first consensus protocol for geo-distributed RSMs that maintains low normal-case latency while tolerating a single fail-slow replica. Avicenna uses a single leader to order commands, naturally tolerating a fail-slow follower. To tolerate a fail-slow leader, Avicenna compares the current latency with the counterfactual latency clients would experience if a different replica, the shadow leader, were the leader. When that comparison indicates the current leader might be slow, Avicenna quickly promotes the shadow leader with a fast leader rotation protocol. Our evaluation shows Avicenna has the same normal-case latency as Multi-Paxos while tolerating fail-slow faults. Reading List: [ https://docs.google.com/document/d/1tSJQW1Z5HbBKJgwAZj09oknyg1-nWeyidS04n4-… | https://docs.google.com/document/d/1tSJQW1Z5HbBKJgwAZj09oknyg1-nWeyidS04n4-… ]

1 0

Raj Ghugare will present his General Exam "Normalizing Flows are Capable Models for RL" on Thursday, April 30, 2026 at 3:00 PM in CS 105.
by CS Grad Department 29 Apr '26

29 Apr '26

Raj Ghugare will present his General Exam "Normalizing Flows are Capable Models for RL" on Thursday, April 30, 2026 at 3:00 PM in CS 105. Committee Members: Benjamin Eysenbach (advisor), Karthik Narasimhan, Brenden Lake Abstract: Modern reinforcement learning (RL) algorithms have found success by using powerful probabilistic models, such as transformers, energy-based models, and diffusion/flow-based models. To this end, RL researchers often choose to pay the price of accommodating these models into their algorithms – diffusion models are expressive, but are computationally intensive due to their reliance on solving differential equations, while autoregressive transformer models are scalable but typically require learning discrete representations. Normalizing flows (NFs), by contrast, seem to provide an appealing alternative, as they enable likelihoods and sampling without solving differential equations or autoregressive architectures. However, their potential in RL has received limited attention, partly due to the prevailing belief that normalizing flows lack sufficient expressivity. We show that this is not the case. Building on recent work in NFs, we propose a single NF architecture which integrates seamlessly into RL algorithms, serving as a policy, Q-function, and occupancy measure. Our approach leads to much simpler algorithms, which often achieve higher performance in imitation learning, offline, goal conditioned RL and unsupervised RL. Reading List: https://docs.google.com/document/d/1cMCRPwSPhpgrE_Qnku0INV3z7dVFFrJeThvF0NZ… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

2 1

Jennifer Lam will present her FPO "Rethinking Distributed Systems for High Contention Workloads" on Monday, 5/4/26 at 10am in CS 301
by Gradinfo 28 Apr '26

28 Apr '26

Jennifer Lam will present her FPO "Rethinking Distributed Systems for High Contention Workloads" on Monday, 5/4/26 at 10am in CS 301 and Zoom Zoom link: [ https://princeton.zoom.us/j/91095935061?pwd=heXLEXo7YrQLfGSdK8ZSVN6r5zv3AF.… | https://princeton.zoom.us/j/91095935061?pwd=heXLEXo7YrQLfGSdK8ZSVN6r5zv3AF.… ] Committee: Examiners: Wyatt Lloyd (adviser), Mae Milano, and Jialin Ding, Zachary Kincaid Readers: Wyatt Llyod and Haonan Lu (co-adviser, external at SUNY Buffalo) Abstract: Strongly consistent distributed systems support large-scale applications by sharding data across many machines to provide capacity greater than what can fit on a single machine. They are designed to run on asynchronous datacenter networks, where transaction requests may be dropped, delayed, or interleaved across shards. Interleaving requests is a root cause of performance bottlenecks. When transactions overlap across shards, their requests arrive in opposite orders. If at least one request is a write, it becomes challenging for serializable systems to guarantee that transactions appear in a matching total order across all shards; writes make the execution order non-commutative, so a read ordered before a write cannot observe it, while a read ordered after must. Consequently, interleaving may result in a request order on one shard contradicting another. Such conflicts degrade performance by consuming extra CPU, memory, and network round-trips. The frequency of conflicts increases under workload skew, where access patterns favor a subset of keys. Skew causes overlap between transactions accessing the same shards and, combined with writes, leads to high contention. Consequently, concurrency control protocols that performed well under low contention often perform poorly under high contention. Many use conflict-based ordering, which orders overlapping transactions using conflict resolution mechanisms such as wait-wait or wait-die. These mechanisms typically allow at most one transaction to proceed at a time, forcing others to wait or be aborted. Under high contention, this worsens overall latency and throughput. This dissertation recognizes that high-contention workloads are challenging for existing distributed systems and calls for new architectures robust against them. We introduce two such systems. The first is TurboDB, a process-ordered serializable distributed database that mitigates conflict on the hottest keys by co-locating them on a hotshard and efficiently serializing requests to that shard. It improves performance by up to an order of magnitude. The second is Whiptail, a strictly serializable distributed database that eliminates conflicts. It exploits improved datacenter networks and tightly synchronized clocks to enable clients to predetermine the total order of transactions, eliminating conflicts. Whiptail achieves up to an order-of-magnitude improvement in throughput on high-contention workloads.

1 0

Arpon Basu will present his General Exam "Sparsifying Sums of Positive Semidefinite Matrices"on April 30, 2026 at 12pm in CS 402.
by Gradinfo 27 Apr '26

27 Apr '26

Arpon Basu will present his General Exam "Sparsifying Sums of Positive Semidefinite Matrices"on April 30, 2026 at 12pm in CS 402. The members of his committee are as follows: Pravesh Kothari (adviser), Zeev Dvir, and Ran Raz Link to Abstract and Reading List: [ https://docs.google.com/document/d/1Ue2YLJakw_u57pi0Fqa5uW57-DNx_GTV0bgifJD… | https://docs.google.com/document/d/1Ue2YLJakw_u57pi0Fqa5uW57-DNx_GTV0bgifJD… ] 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

Xuyang Cao will present his General Exam "Exploring a 5G Architecture for Efficient, QoE-Aware Video Conferencing" on May 4th, 2026 at 12:30pm in Friend Center 005 and Zoom.
by Gradinfo 27 Apr '26

27 Apr '26

Xuyang Cao will present his General exam "Exploring a 5G Architecture for Efficient, QoE-Aware Video Conferencing" on May 4th, 2026 at 12:30pm in Friend Center 005 and Zoom. Zoom Link: [ https://princeton.zoom.us/j/7013752837 | https://princeton.zoom.us/j/7013752837 ] Committee: Kyle Jamieson (adviser), Wyatt Lloyd, Jialin Ding Link to Abstract and Reading List: [ https://docs.google.com/document/d/1HH4njRWhRznymhE3xvzYRwWzwokmKZCWefbC2yP… | https://docs.google.com/document/d/1HH4njRWhRznymhE3xvzYRwWzwokmKZCWefbC2yP… ]

1 0

Sowmya Thanvantri will present her General Exam "Generalized Combinations of Experts in Multi-Head Attention" on Monday, April 27, 2026 at 1:30 PM in CS 401.
by CS Grad Department 24 Apr '26

24 Apr '26

Sowmya Thanvantri will present her General Exam "Generalized Combinations of Experts in Multi-Head Attention" on Monday, April 27, 2026 at 1:30 PM in CS 401. Committee Members: Ryan Adams (advisor), Tom Griffiths, Karthik Narasimhan Abstract: The success of large language models today is attributed largely to the attention mechanism used in transformers, which assigns a measure of uncertainty over other tokens. This idea was extended to multi-head attention, allowing models to learn several different distributions of uncertainty over tokens. However, multi-head attention does not take full advantage of the different distributions because the outputs of the heads are concatenated and passed through a linear map, which prevents learning more expressive mappings. To enhance expressivity, a mixture model could aggregate information between these distributions. Alternatively, taking a Bayesian perspective, a product-based model across distributions could lead to a more accurate representation. To achieve both of these, we draw on ideas from mixture of experts (MoE) and product of experts (PoE) and propose a generalized combination of expert heads in multi-head attention to allow models to learn a richer representation of attention. Reading List: https://docs.google.com/document/d/1PaOPsZcP-rYhBvUYQBqSR-c_BAB3HIYI4rGCSE7… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

2 1

Jiatong Yu will present her MSE thesis "On the Impossibility of Retrain Equivalence in Machine Unlearning" on Friday, April 24th, 2026 in CS 301 at 10:00am.
by gradinfo＠cs.princeton.edu 24 Apr '26

24 Apr '26

Jiatong Yu will present her MSE thesis "On the Impossibility of Retrain Equivalence in Machine Unlearning" on Friday, April 24th, 2026 in CS 301 at 10:00am. Advisor: Sanjeev Arora Reader: Kai Li All are welcome to attend. Abstract: {\em Machine unlearning} seeks to selectively remove the ``influence'' of specific training data on a model's outputs. The ideal goal is {\em Retrain Equivalence}---behavior identical to a model trained from scratch on only the retained data. This goal was originally formulated for models trained on {\em i.i.d.}\ data batches, but modern pipelines often involve multi-stage training, with each stage having a distinct data distribution and objective. Examples include LLM finetuning for alignment, reasoning ability, etc. Building on prior work that has shown retrain-equivalent behavior can be ill-defined due to forgeable datasets in mini-batch training, we study a complementary source of fragility that arises in multi-stage training. In an overparameterized linear regression model, we prove an exponential lower bound on {\em path-dependent divergence} under a family of unlearning objectives, unifying purely local gradient ascent and weakly local algorithms with retain-set regularization. Models trained on the same stages in different orders can be driven exponentially far apart during unlearning, and we show that this pointwise divergence implies violation of the distributional $(\varepsilon,\delta)$-certified unlearning definition. We then empirically measure the same kind of path dependence in LLM post-training across Llama and Qwen models (1B--14B) with gradient ascent, NPO, and SimNPO algorithms.Models finetuned via different orderings of identical training stages diverge in behavior during unlearning, with the degradation in GSM8K accuracy after unlearning varying by over $20\%$ across paths. We also observe that some learning paths consistently produce models that unlearn slowly while preserving higher retained utility---a phenomenon we term the \emph{recency effect}---and that the fate of probability mass during unlearning (e.g., paraphrasing vs.\ alternative concepts) is likewise path-dependent. Taken together, these results quantitatively demonstrate that local, path-oblivious unlearning in multi-stage pipelines can be highly sensitive to training history, adding a multi-stage perspective to the impossibility results and motivating the search for new unlearning desiderata and evaluations.

1 0

Nanqinqin Li will present his FPO "System Idle Time Need Not Be Wasted" on Thursday, 5/7/2026 at 11:30 AM at 194 Nassau Street conference room.
by Gradinfo 23 Apr '26

23 Apr '26

Nanqinqin Li will present his FPO "System Idle Time Need Not Be Wasted" on Thursday, 5/7/2026 at 11:30 AM at 194 Nassau Street conference room. Committee: Examiners: Mike Freedman (adviser), Wyatt Lloyd, Asaf Cidon (Columbia University) Readers: Mike Freedman and Jialin Ding Abstract: Modern computer systems routinely lose performance and availability to time spent waiting on blocking events: CPUs stall on memory, services block on storage and networks, and stateful applications wait through conservative failover protocols. This dissertation studies a common question: when a system must wait, how can it safely run other useful work in parallel with that wait? The dissertation develops this theme through two systems projects. The first, Speculative Recovery, targets failover for stateful applications using recovery from disaggregated storage (REDS). REDS is resource efficient because only one instance runs during normal operation, but failover is slow because timeout and recovery run sequentially. Speculative Recovery starts backup recovery as soon as the primary appears unavailable, while letting the primary continue in case it recovers first. The work introduces disk superposition and the super/collapse abstractions, allowing temporary divergence of disk state while ensuring only one version becomes externally observable. The design includes collocated-clone for near-normal clone performance and dirty-bit-based rules for correctness. Implemented in Ceph and evaluated with MySQL, PostgreSQL, and MariaDB, the approach improves failover while preserving the resource efficiency of REDS. The second project, LiteSwitch, targets sub-microsecond CPU stall cycles caused by CXL-attached memory. CXL expands memory capacity but increases access latency and amplifies memory-induced stalls. Existing harvesting techniques are mismatched: profiling-based methods struggle with CXL latency variation, and interruptbased delivery is too expensive for hundreds-of-nanoseconds windows. LiteSwitch uses a lightweight hardware-software co-design. On the hardware side, location-dependent memory branching (LDMB) detects long-latency accesses online and delivers control via direct branching. On the software side, Bundled Handoff provides fast scavenger selection, and xstate-aware context switching avoids unnecessary SIMD/FP iiisave/restore overhead. Evaluation shows substantial slowdown reductions across representative workloads and CXL latency settings. Taken together, these projects show that idle time can be an opportunity rather than unavoidable loss. The central lesson is that useful parallelization with waiting is effective only when systems codesign performance mechanisms with correctness constraints. By combining overlap with careful control over observability, ordering, and runtime overhead, this dissertation demonstrates practical ways to improve both availability and performance in modern systems.

1 0