- talks - lists.cs.princeton.edu

Bhargav Reddy Godala will present his FPO "Criticality-Aware Front-end" on Tuesday, 5/14/2024 at 11am in CS 402.
by Gradinfo 09 May '24

09 May '24

Bhargav Reddy Godala will present his FPO "Criticality-Aware Front-end" on Tuesday, 5/14/2024 at 11am in CS 402. The members of his committee are as follows: Examiners: David August (adviser), Margaret Martonosi, and David Wentzlaff Readers: Gilles A. Pokam (Intel), Svilen Kanev (Google), and David August Please see abstract below. All are welcome to attend. Code footprints continue to grow faster than instruction caches, putting additional pressure on existing front-end structures. Even with aggressive front-ends with fetch-directed instruction prefetching (FDIP), modern processors experience signifcant front-end stalls. Due to the end of Moore’s Law, increasing cache sizes raises critical path latency, with modest returns for scaling instruction cache sizes. This dissertation aims to address front-end bottlenecks by making two key observations. In FDIP-enabled processors, cache misses have unequal costs, and a small fraction of critical instruction cache lines contribute to most of front-end stalls. EMISSARY, the pioneering cost-aware replacement policy tailored for the L1 Instruction Cache (L1I), defes conventional wisdom by presenting a groundbreaking approach. Unlike traditional replacements, EMISSARY demonstrates performance enhancements even amidst increased instruction cache misses. However, EMISSARY proves to be less efective when applied to datacenter workloads characterized by large code footprints. This is due to datacenter workloads having more critical lines greater than the capacity of L1I. This dissertation frst presents improved EMISSARY-L2, the frst criticality-aware cache replacement family of policies specifcally designed for datacenter workloads. Observing that modern architectures entirely tolerate many instruction cache misses, EMISSARY-L2 resists evicting those cache lines whose misses cause costly decode starvations from L2. In the context of a modern FDIP-enabled processor, EMISSARY-L2 delivers an impressive 3.24% geomean speedup (up to 23.7%) and a geomean energy savings of 2.1% (up to 17.7%) when evaluated on datacenter workloads. This speedup is 21.6% of the speedup obtained by an unrealizable L2 cache with a zero-cycle miss latency for all capacity and confict instruction misses. This dissertation then proposes Priority Directed Instruction Prefetching (PDIP), a novel cost-ware instruction prefetching technique that complements FDIP by issuing iii prefetches for targets along the resteer path where FDIP stalls occur. PDIP identifes these targets and associates them with a trigger for future prefetch. When paired with EMISSARY-L2, PDIP achieves a geomean IPC speedup of 3.7% across a set of datacenter workloads using a budget of only 43.5KB. PDIP achieves 62% of the ideal prefetching performance.

2 1

Kaiqu Liang will present his General Exam "Towards human-centered AI safety" on Monday, May 13, 2024 at 1:00 PM in CS 105.
by CS Grad Department 08 May '24

08 May '24

Kaiqu Liang will present his General Exam "Towards human-centered AI safety" on Monday, May 13, 2024 at 1:00 PM in CS 105. Committee Members: Jaime Fernández Fisac (advisor), Tom Griffiths, Benjamin Eysenbach Abstract: In the domain of human-AI interactions, ensuring the safety of AI systems is crucial, particularly as they are deployed in high-stakes scenarios. This extends beyond the conventional focus on aligning AI objectives with human values to include a deep understanding and respect for human needs as hard constraints within the AI's decision-making process. Our research aims to develop safety-critical human-AI systems to prevent violations of critical human needs proactively. To achieve this, my PhD research actively pursues three critical dimensions: developing safe representations, enhancing human-AI alignment, and advancing safety-aware cognition. In my early Ph.D. research, I took the first step in safety-aware cognition, which quantifies safety risks and uncertainty in the domain of the robot task execution with a Large language models (LLMs) planner. We introduced introspective planning, a novel method enabling robots to proactively identify and seek clarification on uncertainties, effectively minimizing risk without fine-tuning. Through rigorous testing, we demonstrated that introspective planning outperforms existing LLM-based methods in both success rate and safety, showing significant promise for enhancing human-AI interactions in safety-critical settings. This approach not only addresses the direct issue of ensuring compliance and safety in robot task execution but also contributes to the broader objective of developing AI systems that prioritize human needs and safety as foundational principles. Reading List: https://docs.google.com/document/d/14BHiLF8wsrXU2YwErddQ6RD2LKNNIIQeajkHkkg… 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

NextG Seminar: Alberto Valdes-Garcia, Tuesday May 14 at 12:30pm
by Emily C. Lawrence 07 May '24

07 May '24

Speaker: Alberto Valdes-Garcia, IBM T. J. Watson Research Center Title: 3D millimeter-wave imaging and sensing with Si-based phased arrays, edge computing, and AI Day: May 14, 2024 Time: 12:30pm Location: B205, EQuad Hosts: Yasaman Ghasempour and Kaushik Sengupta Event Page: [ https://ece.princeton.edu/events/3d-millimeter-wave-imaging-and-sensing-si-… | https://ece.princeton.edu/events/3d-millimeter-wave-imaging-and-sensing-si-… ] Title : 3D millimeter-wave imaging and sensing with Si-based phased arrays, edge computing, and AI Abstract: The use of millimeter-wave frequencies for 5G networks has been a primary contributor for transitioning Si-based phased array technology from R&D to real-world deployments. While the commercial use of millimeter-wave sensing so far has been dominated by low-cost, compact MIMO radars for automotive and industrial applications, the on-going wide deployment and advancement of Si-based phased arrays opens a new horizon of opportunities for sensing and event recognition. This talk will first cover the fundamentals and KPIs of 3D radar systems using phased arrays including associated key circuit design and packaging design techniques. Examples of such 3D radar systems at 28-GHz, 60-GHz and 94-GHz will be provided. Next, the presentation will describe how the full potential of such systems can be realized through synergistic co-design with algorithms and edge computing assets. Key examples of emerging applications based on these vertically integrated antennas-to-software/AI systems will be provided including multi-spectral imaging, 5G mmWave joint sensing and communications, and AI-based recognition of human gestures and concealed objects. Bio: Alberto Valdes-Garcia is currently a Principal Research Scientist and Manager of the RF Circuits and Systems Group at the IBM T. J. Watson Research Center. In his current role, he leads a multi-disciplinary team that investigates and develops technologies that bridge the gap between antennas and edge-compute-based AI, enabling new millimeter-wave systems and applications for imaging and communications. Dr. Valdes-Garcia received the Ph.D. degree in Electrical Engineering from Texas A&M University in 2006. He holds >140 issued US patents and has authored >140 peer-reviewed publications. Recent awards include the 2017 Lewis Winner Award for Outstanding Paper, presented by the IEEE International Solid-State Circuits Conference, and the 2017 IEEE Journal of Solid-State Circuits Best Paper Award. In 2013, he was selected by the National Academy of Engineering for its Frontiers of Engineering Symposium. He currently serves in the Inaugural Editorial Board of the IEEE Journal of Microwaves and was the Chair of the IEEE MTT-S Microwave and Millimeter-Wave Integrated Circuits Committee in 2020-2021. Dr. Valdes-Garcia is a Fellow of IEEE, was inducted into the IBM Academy of Technology in 2015, and was recognized as an IBM Master Inventor in 2016, 2019, and 2022.

1 0

PPH Lunch seminar: Nir Eyal, Friday 5/10 at 12 pm @ 252 Nassau St
by Emily C. Lawrence 07 May '24

07 May '24

Please join the Princeton Precision Health Initiative for the next talk in our lunch time seminar series this Friday, May 10, at 12:00 pm, in our new space at 252 Nassau Street. Nir Eyal, the founding director of the Rutgers Center for Population-level Bioethics in the Rutgers University School of Health will present a talk titled, “The ethics of human challenge trials for emerging disease outbreaks”. Lunch will be provided. Please note that 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 ] The PPH Team ************************************ Princeton Precision Health [ http://pph.princeton.edu/ | pph.princeton.edu ] [ mailto:princetonpph@princeton.edu | princetonpph(a)princeton.edu ]

1 0

Dmitry Paramonov will present his FPO "Handling Data Too Large To Handle: On Multi-pass Streaming and Interactive Coding" on May 7th, 2024 at 11am in Friend 006.
by Gradinfo 07 May '24

07 May '24

Dmitry Paramonov will present his FPO "Handling Data Too Large To Handle: On Multi-pass Streaming and Interactive Coding" on May 7th, 2024 at 11am in Friend 006. His committee is as follows: Examiners: Gillat Kol (adviser), Ran Raz, Mark Braverman Readers: Huacheng Yu and Matt Weinberg Title: Handling Data Too Large To Handle: On Multi-pass Streaming and Interactive Coding Abstract: Over the last decades, the world has become increasingly more information-centric and massive amounts of data, potentially distributed between many different sources, are being processed all the time. In this thesis, I consider two mechanisms for coping with big data and the distributed nature of timely tasks. In Part I, I showcase my work on the streaming setting, where the input to the algorithm is given as a stream of elements. The algorithm’s goal is to compute a value that depends on the stream while only utilizing memory that is much smaller than the entire stream. My work in this field focuses on proving that various fundamental graph problems essentially require the streaming algorithm to store the entire graph, even if it is allowed to make several passes through the given stream of edges. In Part II, I consider error-correcting codes for distributed, interactive settings. Classical error- correcting codes assume that a sender who has all the information wishes to send it to a receiver over a noisy channel. However, in many modern, big data applications, the information is distributed amongst many parties that communicate back-and-forth to compute a value that depends on all their inputs. My work examines the noise resilience of various such settings. For some models, we can design error-correcting protocols that allow the encoding of every noiseless protocol by a noise-resilient protocol with low overhead, whereas for other models, it can be shown that this task is impossible. While these two topics appear greatly unrelated, and almost orthogonal to one another, the tools used to prove results in both turn out to be remarkably similar, with many standard problems and information theory lemmas being a critical part of both.

1 1

Chenyue Cai will present her General Exam "Geometry and Directional Aware Deep Detail Transfer from a Single 3D Model." on Thursday, May 9, 2024 at 3:30 PM in CS 301 and via zoom.
by CS Grad Department 06 May '24

06 May '24

Chenyue Cai will present her General Exam " Geometry and Directional Aware Deep Detail Transfer from a Single 3D Model. " on Thursday, May 9, 2024 at 3:30 PM in CS 301 and via zoom. Zoom link: https://princeton.zoom.us/j/95829239477 Committee Members: Szymon Rusinkiewicz (advisor), Felix Heide , Adam Finklestein Abstract: As the demand for 3D content creation is on the rise, it is still notoriously time-consuming for artists to create high quality, detailed 3D models. One significant challenge lies in the intricate detailing of 3D shapes. Traditional synthesis methods generate details procedurally, patch by patch, and then blend them with the underlying geometry or texture. More recently, 3D texture generation methods using pre-trained image generation have shown promise, but they often require a large training set and struggle to produce high-quality details in the generated assets due to view inconsistency, occlusion and distortion. In this talk, I will introduce my work in transferring details from one single high quality model to arbitrary shapes. I will talk about the algorithm that aims to tackle the two main challenges in texture synthesis on 3D shape: first is the need for the transfer to respond to local geometry, second is to make the texture anisotropic and directional aligned. The algorithm first uses a geometric feature extraction module to develop a probabilistic matching of source and target surface. Then, it generates paired directional fields which will later guide the synthesis process. Lastly, the algorithm directly synthesizes texture on surface by mapping geometric features to the learned texture space through a deep network guided by the directional field. I will also discuss the design choices and challenges inherent in this approach, as well as its potential applications and future developments. Reading List: https://docs.google.com/document/d/1HyS0tGw6ZmqgJrrpb6VJ5oKkZZXnTBrLI5ENFlg… 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

Yuhan Liu will present her General Exam "Understanding and Building Decentralized Sociotechnical Systems" on Wednesday, May 8, 2024 at 8:00 AM in CS 402 and via zoom.
by CS Grad Department 06 May '24

06 May '24

Yuhan Liu will present her General Exam "Understanding and Building Decentralized Sociotechnical Systems" on Wednesday, May 8, 2024 at 8:00 AM in CS 402 and via zoom. Zoom link: https://princeton.zoom.us/j/4652800681 Committee Members: Andrés Monroy-Hernández (advisor), Parastoo Abtahi, Janet Vertesi Abstract: Since the 19th century, the impact of technological advancements on labor has been recognized as a double-edged sword. The rise of scientific management, also known as Taylorism, increased production efficiency but also led to worker exploitation. Today, the evolution of algorithm-mediated labor reflects this dynamic, with the arena shifting from factories to different mobile applications and the source of stress evolving from timers on the assembly lines to AI-powered algorithms. The past decade has witnessed the rise of the gig economy, alongside significant challenges. For example, monopolistic food delivery platforms are the sole authority in deciding restaurant commissions, customer prices, and worker wages, causing imbalances of power. The black-boxed AI algorithms result in significant information asymmetry and the gig economy’s zero-sum nature leads to value misalignments between platforms and other stakeholders. My work aims to explore and develop decentralized solutions, drawing inspiration from the labor movement a hundred years ago and the recent decentralization trends in social media. To understand existing alternatives to mainstream gig platforms, I led a mixed-method study investigating the widespread but underresearched socio-technical phenomena: indie food delivery platforms. We collected data on 495 indie platforms and detailed survey responses from 29 platforms. We analyzed these platforms to understand why they were created, how they operate, what technologies they use, and what challenges they face. Furthermore, to dig deeper into the recent trend of migration to decentralized social media, I led a qualitative study on understanding how the nature of decentralization in systems like Mastodon, is associated with novel uses of the technology. Building on the insights from independent food delivery platforms and decentralized social media, I’m now working on designing a protocol for the decentralized food delivery network: Open Deli Protocol. In my future research, we plan to implement and deploy the decentralized food delivery system with the support of our industrial collaborators. The system will improve the user experience and agency of delivery workers and enable indie food delivery platforms to operate in a way that aligns with their values. Reading List: https://docs.google.com/document/d/1vSxL9e_Z6_UfVuahVfzG0sJoZ_BdKC7mGyAHYwm… 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

DeCenter seminar: Ben Fisch, May 8th at 4:30pm
by Emily C. Lawrence 06 May '24

06 May '24

The DeCenter is delighted to extend an invitation to a seminar featuring Ben Fisch on May 8th at 4:30pm . The seminar will be held in the Computer Science Building room 105 , followed by a light reception. Talk Title: How Cryptographic Proofs are Transforming the Decentralized Web Abstract: Ethereum, along with a growing number of decentralized virtual machines, hosts thousands of applications (dapps) that are updated synchronously and secured by a common consensus protocol. These characteristics have made these platforms particularly suitable for financial use cases, where users move funds in and out of dapps for trading, lending, payments, liquidity provision, and more. Users of Ethereum can interact with many dapps at once, such as receiving a loan and executing a trade in a single transaction (aka flash loan), without needing to trust anyone beyond Ethereum’s consensus for the safety of their funds. This shared infrastructure across dapps, however, has several downsides. Ethereum’s fixed transaction throughput creates a scaling bottleneck. An activity spike in one application can cause congestion, raising transaction prices in all others. Dapps are also forced to use Ethereum’s limited instruction set. Operations (like hash functions) that may be cheaply executed on most modern machines can blow up in complexity post compilation to Ethereum instructions. Lastly, by controlling the ordering of transactions, Ethereum’s operators profit through frontrunning or arbitrage — behavior which dapp creators may either want to prevent entirely or capture value from on their own. Today, cryptographic proof systems are enabling dapps to address these issues by moving their transaction processing off Ethereum while still inheriting Ethereum’s security. These are called “Layer-2” applications. Dedicated dapp servers (called “rollup sequencers”) collect, order, and execute large batches of transactions on their own, periodically posting only the resulting state change and a cryptographic proof of its correctness to Ethereum for verification. While this may limit users’ ability to interact synchronously with multiple dapps (e.g., for a flash loan), further reductions in the latency and cost of cryptographic proofs will re-enable atomic cross-dapp transactions. This new paradigm has also led to new mechanism design challenges, such as how to optimize and share cross-dapp revenue, or how to outsource cryptographic proof generation and appropriately price transactions to cover these costs. This talk will be a general presentation of this evolution and discussion of open challenges. Bio: Ben Fisch is an assistant professor of computer science at Yale University and co-director of the Yale Applied Cryptography Lab. His research centers around cryptographic proofs and decentralized systems, but broadly spans topics within privacy and verifiability. Prior to Yale, he completed his PhD at Stanford University. His thesis work included Proofs of Replication and Verifiable Delay Functions, which were influential in the designs of real-world decentralized systems, including Filecoin, Chia, and Ethereum. More recently, he co-founded Espresso Systems, which is improving the decentralization, security, and interoperability of “Layer-2” applications.

1 0

Xinran Liang will present her General Exam "Reducing Contextual Bias using Synthetic Augmentations from Generative Models" on Wednesday, May 8, 2024 at 3:00 PM in CS 302.
by CS Grad Department 03 May '24

03 May '24

Xinran Liang will present her General Exam "Reducing Contextual Bias using Synthetic Augmentations from Generative Models" on Wednesday, May 8, 2024 at 3:00 PM in CS 302. Committee Members: Olga Russakovsky (advisor), Szymon Rusinkiewicz, Jia Deng Abstract: Many vision datasets are known to contain contextual bias. For example, training images of “skateboard” very usually co-occurs with “person”, while in rare cases skateboards occur exclusively. Models trained on these biased datasets tend to leverage co-occurrence between objects and their contexts to improve recognition accuracy. Consequently these models often fail to generalize to scenarios where such co-occurrence patterns are absent. Previous works addressed this problem by enforcing models to decorrelate learned features of a category from its co-occurring context. In this project, with recent advances in generative models, we explore the possibility of using synthetic samples from generative models to reduce contextual bias. We generate useful variations of training data that do not contain typical co-occurring patterns and use them as augmentations in training downstream classifiers. We evaluate our framework on object classification tasks and show that it improves downstream performance compared to standard classifiers. We hope our work could demonstrate the effectiveness of using generated samples to improve limitations in vision datasets and learn better vision models, potentially overcoming the challenge of manually collecting and annotating curated data at scale. Reading List: https://docs.google.com/document/d/1YWZjPo272HKOp9CLd6v4yK1RzXejGr4mit12lcY… 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

Dillon Lue will present his General Exam "Multimodal Single Cell Genomics Datasets for Identifying Cell Types and their Function" on Wednesday, May 8, 2024 at 1:30 PM in Icahn 280.
by CS Grad Department 03 May '24

03 May '24

Dillon Lue will present his General Exam "Multimodal Single Cell Genomics Datasets for Identifying Cell Types and their Function" on Wednesday, May 8, 2024 at 1:30 PM in Icahn 280. Committee Members: Yuri Pritykin (advisor), Ben Raphael, Olga Troyanskaya Abstract: Recent advances in single-cell genomics enable the simultaneous measurement of RNA and another modality in the same cell, known as multimodal scRNA-seq. In this talk, I demonstrate how multimodal assays can enhance our understanding of T cell biology and inform scRNA-seq analysis. First, I will talk about the analysis of a multimodal scRNA-seq assay with a novel combination of modalities: barcoded lineage tracing, CRISPR perturbations, and RNA. This combination of modalities enables us to interrogate how CRISPR perturbations influence the transcriptional states in T cell clones. Next, I propose a new framework to benchmark scRNA-seq analysis by utilizing the second modality in multimodal datasets as ground truth. Due to the complex nature of scRNA-seq data, a multi-step procedure has emerged to map cells onto a low dimensional transcriptional manifold. This procedure has many user-defined choices that can influence biological results. Unlike existing benchmarks, this new framework introduces a ground truth for the context of continuous biological processes. Reading List: https://docs.google.com/document/d/12K-XVW5NfBQZ5MBXesUvJlqHBIoA9blOhI0CYHJ… 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