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

Fan Yi will present his FPO "Cross-Layer Telemetry and Optimization for Real-Time Interactive Applications in 5G Networks" on Tuesday, July 8, 2025 at 3pm in Friend 007.
by Gradinfo 30 Jun '25

30 Jun '25

Fan Yi will present his FPO "Cross-Layer Telemetry and Optimization for Real-Time Interactive Applications in 5G Networks" on Tuesday, July 8, 2025 at 3pm in Friend 007. The members of his committee are as follows: Examiners: Kyle Jamieson (adviser), Ravi Netravali, Maria Apostolaki (ECE) Readers: Wyatt Lloyd, Yaxiong Xie (University at Buffalo, SUNY) All are welcome to attend. Abstract: Real-time applications such as interactive video conferencing now underpin daily communication, yet their performance over 5G networks is often inconsistent and unreliable. While 5G promises unprecedented speed, its internal complexity introduces significant fluctuations in latency and throughput that directly degrade the user's Quality of Experience (QoE). These performance issues persist because the network stack operates in functional silos, obscuring the complex, cross-layer interactions that are the true root cause of application-level degradation. The fundamental challenge lies in the highly dynamic wireless channel and the lack of cross-layer visibility. Dynamic channel conditions and millisecond-level scheduling at the physical and link layers continually reshape available capacity, yet application-layer congestion controllers perceive only averaged, coarse-grained feedback and therefore misinterpret the wireless artifacts as generic network congestion. This flawed perception causes them to overreact by unnecessarily throttling sending rates or increasing buffer delays, which impairs the user experience. This thesis tackles this challenge by building a suite of intelligent wireless systems that first expose millisecond-level cross-layer telemetry, and then exploit it for real-time 5G channel optimization and sensing. This work first introduces Athena, a novel measurement framework that correlates high-resolution telemetry from the 5G physical layer up to the video conferencing application. Building on this framework, Domino is developed as a first-of-its-kind diagnostic tool that automates the root-cause analysis of QoE degradation by modeling and searching a graph of causal relationships, successfully identifying 24 previously unknown event chains that impair 5G video conferencing. Moving from diagnosis to direct network enhancement, WaveFlex is the first smart surface for private CBRS 5G that monitors control channels and steers signals in real time, increasing median throughput by 19.5%. Finally, CAPER repurposes cellular measurements for privacy-preserving proximity sensing, achieving highly accurate proximity detection via a custom deep learning pipeline. Collectively, these contributions deliver a comprehensive paradigm for building intelligent wireless systems. By providing the essential tools for deep measurement (Athena), automated diagnostics (Domino), proactive network optimization (WaveFlex), and novel sensing capabilities (CAPER), this thesis establishes the foundational visibility and control necessary to unlock the full potential of 5G and future wireless networks.

1 0

PPH Spring Seminar- Prof. Zohar Yakhini, Monday, June 23rd at 12pm at 252 Nassau St
by Emily C. Lawrence 20 Jun '25

20 Jun '25

Reminder! Please join the Princeton Precision Health (PPH) Initiative for the next talk of our spring semester series on Monday, June 23rd, at 12:00 pm, at 252 Nassau Street. Prof. Zohar Yakhini, from the Efi Arazi School of Computer Science and Head of Machine Learning and Data Science Program, at Reichman University, will present virtually, “Statistical enrichment, image analysis and spatial biology”. The spatial context of molecular and morphological characteristics of cells and cell populations can be very important in determining pathogenesis or other higher level biological processes. It is therefore desirable to understand molecular profiles in a spatial context. In this talk I will describe the application of computational and statistical techniques - ranging from assessing enrichment in ranked lists to applying deep learning-based image analysis - to various spatial/molecular measurement modalities. Modalities addressed include H&E imaging, IHC and spatial transcriptomics. I will describe methods, and some example results and will discuss the potential impact in terms of life science and medical science. Lunch will be provided. Getting to the seminar space currently requires that you climb a set of stairs. If an accommodation is needed, please contact PPH in advance at: [ mailto:PrincetonPPH@princeton.edu | PrincetonPPH(a)princeton.edu ] Thank you, Princeton Precision Health

1 0

Anunay Kulshrestha will present his FPO "Advancing Public Policy Through Cryptography: Building Private and Accountable Systems" on Friday, June 20, 2025 at 11am in Sherrerd 308 (CITP Conference Room) and Zoom.
by Gradinfo 18 Jun '25

18 Jun '25

Anunay Kulshrestha will present his FPO "Advancing Public Policy Through Cryptography: Building Private and Accountable Systems" on Friday, June 20, 2025 at 11am in Sherrerd 308 (CITP Conference Room) and Zoom. Zoom: https://princeton.zoom.us/j/94232769862 All are welcome to join. Examiners: Jonathan Mayer (Adviser), Prateek Mittal, Aleksandra Korolova Readers: Jennifer Rexford, Seny Kamara (Brown University), and Kobbi Nissim (Georgetown University) Abstract: We illustrate how modern cryptography can offer novel solutions to longstanding policy challenges improving both privacy and accountability—–objectives that are often in tension with each other—by discussing examples from content moderation and foreign intelligence surveillance. We formalize the relevant functionalities and design practical protocols that realize them. We provide security proofs, open-source implementations, and benchmarks to demonstrate feasibility of our protocols. We describe limitations of technical answers to complex socio-technical questions. We conclude by discussing the policy impact of our contributions and promising avenues for future work.

1 1

Please Join us at 252 Nassau Street, on June 23rd at 12 pm for our PPH Spring Seminar- Prof. Zohar Yakhini will present virtually “Statistical enrichment, image analysis and spatial biology”.
by Emily C. Lawrence 17 Jun '25

17 Jun '25

Please join the Princeton Precision Health (PPH) Initiative for the next talk of our spring semester series on Monday, June 23rd, at 12:00 pm, at 252 Nassau Street. Prof. Zohar Yakhini, from the Efi Arazi School of Computer Science and Head of Machine Learning and Data Science Program, at Reichman University, will present virtually, “Statistical enrichment, image analysis and spatial biology”. The spatial context of molecular and morphological characteristics of cells and cell populations can be very important in determining pathogenesis or other higher level biological processes. It is therefore desirable to understand molecular profiles in a spatial context. In this talk I will describe the application of computational and statistical techniques - ranging from assessing enrichment in ranked lists to applying deep learning-based image analysis - to various spatial/molecular measurement modalities. Modalities addressed include H&E imaging, IHC and spatial transcriptomics. I will describe methods, and some example results and will discuss the potential impact in terms of life science and medical science. Lunch will be provided. Getting to the seminar space currently requires that you climb a set of stairs. If an accommodation is needed, please contact PPH in advance at: [ mailto:PrincetonPPH@princeton.edu | PrincetonPPH(a)princeton.edu ] Thank you, Princeton Precision Health

1 0

Ted Zadouri will present his General Exam "Hardware-Efficient Attention for Inference" on Friday, June 6, 2025 at 1:00 PM via Zoom.
by gradinfo＠cs.princeton.edu 06 Jun '25

06 Jun '25

Ted Zadouri will present his General Exam "Hardware-Efficient Attention for Inference" on Friday, June 6, 2025 at 1:00 PM via Zoom. Zoom link: https://princeton.zoom.us/j/94965926485 <https://www.google.com/url?q=https://princeton.zoom.us/j/94965926485&sa=D&s ource=calendar&ust=1749645969349347&usg=AOvVaw3t8uU_z00Dbt6kPc88ZdtS> Committee Members: Tri Dao (advisor), Kai Li, Chi Jin Abstract: In light of test-time compute, inference efficiency drives progress in AI, demanding a greater emphasis on inference-aware architectures. The sequential nature of decoding limits parallelism. For large batches and long contexts, the key value cache often bottlenecks decoding as it consumes scarce GPU memory, and each decoding step must load the large cache from HBM. Fetching this cache dominates latency relative to the small matrix-vector computation per decoding step, causing prolonged low GPU utilization. This bottleneck hinders a wide range of use cases, including latency-sensitive requests, multi-step reasoning agents, long-context video models, and test-time compute scaling. This work redesigns attention through the lens of arithmetic intensity to shrink the KV cache size, thereby accelerating decoding without sacrificing quality or distributed scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfer while preserving accuracy. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding at full quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2x faster than FlashMLA, e.g., in a speculative decoding setting where the query length exceeds one. Additionally, loading a smaller KV cache per device further reduces latency and boosts throughput in online serving benchmarks. Reading List: https://docs.google.com/document/d/1aNdCzisnXfuicP_7-sqVECLbILa55uwhWha3oXPu GoM/edit?usp=sharing <https://www.google.com/url?q=https://docs.google.com/document/d/1aNdCzisnXf uicP_7-sqVECLbILa55uwhWha3oXPuGoM/edit?usp%3Dsharing&sa=D&source=calendar&us t=1749645969349347&usg=AOvVaw2IiMs_Y0APTE_5i7jhQttW> Everyone is invited to attend the talk, and those facultywishing to remain for the oral exam following are welcome to do so.

1 0

Satadal Sengupta will present his Pre-FPO "Protocol-Aware Telemetry For In-Network Control" on June 11th, 2025 at 12:30pm in Friend 125 and Zoom.
by Gradinfo 04 Jun '25

04 Jun '25

Satadal Sengupta will present his Pre-FPO "Protocol-Aware Telemetry For In-Network Control" on June 11th, 2025 at 12:30pm in Friend 125 and Zoom. Committee: Jennifer Rexford (Advisor) Ravi Netravali (Examiner) Maria Apostolaki (Examiner, Princeton ECE) David Walker (Reader) Hyojoon Kim (Reader, University of Virginia CS) Zoom: https://princeton.zoom.us/j/2332835952?pwd=tYVB2ta8EI4aC3eX4biZCbu4VyYK0J.1 Abstract: The average Internet user consumes a staggering 4 GB of data per day in 2025, compared to just 1 MB per month in 1993 when the web went public. Yet, despite major improvements, the underlying infrastructure still does not guarantee performance or security, and is afflicted by issues such as packet loss, excessive latency, and suboptimal paths. To address these deficiencies, sophisticated network protocols have been developed: for example, TCP aims to provide reliable delivery of on-demand traffic, and WebRTC addresses timely delivery of interactive traffic—together accounting for over 80% of all Internet traffic. These protocols rely on the Border Gateway Protocol for policy-driven routing. Despite the effectiveness of these protocols, significant performance and security problems still persist. Among these problems, short‑lived or stealthy issues are particularly damaging because they often go undetected by traditional coarse‑grained (flow‑level or packet sampling-based) monitoring, yet cause disproportionate impacts on network performance and security. For example, nearly a quarter of business users of video-conferencing applications report lags during peak hours due to transient network congestion, and billions of dollars are lost yearly to stealthy hijacks of cryptocurrency services. Tackling these problems requires protocol‑aware methods to monitor and control the network in real time at a fine granularity (i.e., packet level). This is a challenging task at Internet scale. State-of-the-art networking hardware, such as programmable switches, is characterized by Tbps-scale speed and programmability, offering the only practical opportunity for fine‑grained, real‑time network visibility and control. However, their high speed comes at the cost of limited programmability in the form of stringent memory and computational constraints. This makes it challenging to analyze traffic governed by complex protocols since the hardware cannot retain complete protocol state under such tight resource constraints. In this dissertation, we address these challenges by developing efficient, hardware‑amenable techniques for real-time network monitoring and control. First, we measure a critical packet‑level metric—TCP round‑trip time (RTT)—to enable prompt detection and mitigation of stealthy long‑distance routing attacks. Second, we analyze popular, representative video‑conferencing applications, such as Zoom (proprietary) and MediaSoup (open‑source, WebRTC‑based), and extract detailed packet‑level performance metrics. Our insights drive the design of a scalable, hardware‑software co‑designed architecture for video‑conferencing, significantly improving performance and reducing operational costs compared to conventional software solutions. In summary, this work advances scalable, protocol‑aware techniques for network monitoring and control to enhance the performance and security of the Internet.

1 0

Zhou Lu will present his FPO "Online Learning: Optimization, Control, and Learning Theory" on Monday, June 16, 2025 at 1:00 PM in Friend 005.
by gradinfo＠cs.princeton.edu 03 Jun '25

03 Jun '25

Zhou Lu will present his FPO "Online Learning: Optimization, Control, and Learning Theory" on Monday, June 16, 2025 at 1:00 PM in Friend 005. The members of Zhou's committee are as follows: Examiners: Elad Hazan (Adviser), Chi Jin, Ryan Adams Readers: Matt Weinberg, David Woodruff (CMU) A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of thethesis. Everyone is invited to attend his talk. Abstract follows below: Online learning is a foundational paradigm in learning theory, addressing the challenge of making predictions from sequential data. Initially introduced as a mathematical model of learnability, online learning has since evolved into a versatile framework with widespread applications in optimization, control, economics, and beyond, spurring algorithmic innovations and empirical advances. This dissertation investigates various theoretical aspects of online learning, encompassing both its fundamental limits and its applications to optimization and control. The first part of this dissertation focuses on online convex optimization. We develop more efficient algorithms for adaptive regret minimization, improving both query and projection efficiency. We also address the more challenging setting of online non-convex optimization via a novel reduction approach. The second part of the dissertation is devoted to online nonstochastic control, a generalization of classical optimal control that relaxes assumptions on the cost structure and disturbance models. We propose a new framework for controlling marginally stable linear dynamical systems, a notoriously difficult class of systems. We also design an optimal algorithm for bandit nonstochastic control with general loss functions, and develop a meta-algorithm capable of aggregating multiple base control algorithms. The final part of the dissertation turns to core problems in learning theory. We establish a necessary and sufficient condition for inductive reasoning, a foundational problem in epidemiology, by drawing a new connection to online learning theory. Along the way, we also present new results in multimodal learning theory, building provable statistical and computational advantages. Together, these contributions advance the theoretical understanding of online learning and extend its applicability to a broader range of domains, offering new insights and tools for both researchers and practitioners.

1 0