April 2023 - talks - lists.cs.princeton.edu

Christopher Hodsdon will present his FPO "Stronger Abstractions and Performance Guarantees for Building Strongly Consistent Distributed Services" on May 9th, 2023 at 11am in CS 302.
by Gradinfo 11 Apr '23

11 Apr '23

Christopher Hodsdon will present his FPO "Stronger Abstractions and Performance Guarantees for Building Strongly Consistent Distributed Services" on May 9th, 2023 at 11am in CS 302. His committee is as follows: Examiners: Wyatt Lloyd (advisor), Ethan Katz-Bassett (advisor), Michael Freedman Readers: Siddhartha Sen, Ravi Netravali Everyone is invited to attend the talk. Title: Stronger Abstractions and Performance Guarantees for Building Strongly Consistent Distributed Services Abstract: Strongly consistent distributed services require complex coordination among the services’ machines. Building these services thus requires designing and reasoning about complex coordination. To make building strongly consistent distributed services easier, designers build services atop abstractions that simplify reasoning about the service. Two abstractions used today are the multi-sequence abstraction and replicated state machines. The multi-sequence abstraction simplifies ordering operations on sharded data; it explicitly orders the operations on each shard it accesses. Existing implementations have two shortcomings: failures can result in some multi-sequence numbers never being assigned, exposing a noncontiguous multi-sequence, which requires designing and implementing complex coordination in the form of consensus to handle, and existing implementations do not scale The replicated state machine (RSM) abstraction is used by many of today’s services to ensure correctness and resilience despite failures. It provides the abstraction of “a single machine that does not fail”. However, with RSM protocols used today, an individual slow replica can slow the entire RSM. RSMs are commonly geo-distributed, but current state-of-the-art protocols that provide the stronger slowdown-tolerant abstraction do not have comparable latency to the state-of-practice in a wide-area network or are unable to handle certain types of slowdowns. This dissertation shows that it is possible to build systems that strengthen both abstractions while providing performance guarantees. First, we posit that sequencers should expose our new contiguous multi-sequence abstraction. Contiguity guarantees every sequence number is assigned an operation, simplifying and strengthening the abstraction. Without scalability in the implementation of the contiguous multi-sequence, a service would need to be redesigned once it has outgrown the implementation. We design and implement MASON, the first system to expose the contiguous multi-sequence abstraction and the first to provide a scalable multi-sequence. MASON is thus an ideal building block for consistent, scalable services. Our evaluation shows MASON unlocks scalable throughput for two strongly consistent services built on it. Second, we seek to provide the stronger RSM abstraction of “a single machine that does not slow down”, with AVICENNA, a slowdown-tolerant replicated state machine protocol that has comparable latency to the current state-of-practice in the wide-area.

1 0

Promise Ekpo will present her MSE talk "Investigating Emergent Persuasiveness in Large Language Models" on Wednesday, April 19, 2023 at 12:30pm in CS 302.
by Nicki Mahler 11 Apr '23

11 Apr '23

Promise Ekpo will present her MSE talk "Investigating Emergent Persuasiveness in Large Language Models" on Wednesday, April 19, 2023 at 12:30pm in CS 302. The committee is as follows: Adviser: Prof. Jaime Fernandez Fisac Reader: Prof Karthik Narasimhan All are welcome to attend. Abstract: While the rate of progress and innovation in Artificial intelligence (AI) has many potential benefits, the potential for accidental deleterious effects cannot be overemphasized. It has been empirically demonstrated that large language models (LLMs) can learn to perform a wide range of natural language processing (NLP) tasks in a self-supervised setting. However, these models might unintentionally produce convincing arguments for false statements. There has been recent interest in improving LLM performance by fine-tuning in a reinforcement learning framework through interaction with human users. One could raise the concern that even seemingly benign reward functions can lead to goal-driven manipulation of user responses as an instrumental goal to achieve higher overall performance. This thesis seeks to investigate this possibility by evaluating the persuasiveness of self-supervised-only and reinforcement-learning-fine-tuned LLMs. In this work, we will discuss three approaches to investigating the degree of emergent persuasiveness in LLMs by searching for qualitative failures through a direct query, quantifying the persuasiveness of generated outputs, and training on this persuasiveness metric as a reward signal with reinforcement learning. Through our investigation, we find that state-of-the-art LLMs fail when prompted with statements about less popular misconceptions or domain-specific myths. With this investigation of the safety-critical related failures of LLMs, we hope to further inform the public of the degree of reliability of these models and guide their use.

1 0

Theano Stavrinos will present her FPO "Leopard: Unlocking Better Cache Performance at Lower Cost with Expiration Time-based Flash Caching" on Tuesday, May 9, 2023 at 1pm in CS 302.
by Gradinfo 11 Apr '23

11 Apr '23

Theano Stavrinos will present her FPO "Leopard: Unlocking Better Cache Performance at Lower Cost with Expiration Time-based Flash Caching" on Tuesday, May 9, 2023 at 1pm in CS 302. The committee is as follows: Examiners: Wyatt Lloyd (adviser), Kai Li, and Amit Levy Readers: Ethan Katz-Bassett (Columbia/co-adviser) and Ravi Netravali All are welcome to attend. Title: "Leopard: Unlocking Better Cache Performance at Lower Cost with Expiration Time-based Flash Caching" Abstract: Caches are crucial building blocks of web services. They keep data close to users and other services, reducing request latencies, expensive network traversals, and requests to resource-constrained backend servers. Today’s web services need high-capacity, high-performance caches for their massive working set sizes and to meet stringent performance requirements. Flash-based SSDs meet this need by providing excellent performance and high capacity at low cost. However, caching on flash involves a fundamental tradeoff. On the one hand, caches aim for low miss ratios by keeping useful objects in the cache. On the other hand, caches must protect SSDs from write-induced wear-out, which increases when useful objects are copied forward during garbage collection. Flash caches are often forced to choose between good cache performance (i.e., low cache miss ratios) and acceptable device lifespans. This dissertation describes Leopard, a flash caching framework for static content that unlocks new positions along the Pareto frontier of the miss ratio/device lifespan tradeoff, enabling more effective caching at lower cost than existing frameworks. At the foundation of Leopard's design are expiration times, which specify an object's earliest eviction time. In particular, Leopard’s garbage collection procedure uses expiration times in a cost/benefit analysis to choose the best flash block to erase, balancing write amplification from copying forward useful objects and increased miss ratios from evictions. Leopard also uses a novel clustering algorithm to group objects together by their expiration times, increasing the likelihood of low write amplification during garbage collection. Our evaluation shows that, for a range of CDN traces, Leopard significantly improves the object miss ratio achievable at a given total write volume compared to the state-of-the-art flash caching framework. Leopard also achieves better byte miss ratios at lower write volumes at most points along the byte miss ratio/write volume Pareto frontier.

1 0

ECE Seminar: Andreea Bobu, UC Berkeley, TODAY at 4:30pm
by Emily C. Lawrence 11 Apr '23

11 Apr '23

ECE SEMINAR Speaker: Andreea Bobu, University of California Berkeley Title: Aligning Robot and Human Representations Day: Tuesday, April 11, 2023 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: Chi Jin Abstract: To perform tasks that humans want in the world, robots rely on a representation of salient task features; for example, to hand me a cup of coffee, the robot considers features like efficiency and cup orientation in its behavior. Prior methods try to learn both a representation and a downstream task jointly from data sets of human behavior, but this unfortunately picks up on spurious correlations and results in behaviors that do not generalize. In my view, what's holding us back from successful human-robot interaction is that human and robot representations are often misaligned: for example, our lab's assistive robot moved a cup inches away from my face -- which is technically collision-free behavior -- because it lacked an understanding of personal space. Instead of treating people as static data sources, my key insight is that robots must engage with humans in an interactive process for finding a shared representation for more efficient, transparent, and seamless downstream learning. In this talk, I focus on a divide and conquer approach: explicitly focus human input on teaching robots good representations before using them for learning downstream tasks. This means that instead of relying on inputs designed to teach the representation implicitly, we have the opportunity to design human input that is explicitly targeted at teaching the representation and can do so efficiently. I introduce a new type of representation-specific input that lets the human teach new features, I enable robots to reason about the uncertainty in their current representation and automatically detect misalignment, and I propose a novel human behavior model to learn robust behaviors on top of human-aligned representations. By explicitly tackling representation alignment, I believe we can ultimately achieve seamless interaction with humans where each agent truly grasps why the other behaves the way they do. Bio: Andreea Bobu is a Ph.D. candidate at the University of California Berkeley in the Electrical Engineering and Computer Science Department advised by Professor Anca Dragan. Her work is at the intersection of robotics, machine learning, and mathematical human modeling. Specifically, Andreea studies algorithmic human-robot interaction, with a focus on how autonomous agents and humans can efficiently and interactively arrive at shared representations of their tasks for more seamless and reliable interaction. Prior to her Ph.D. she earned her Bachelor's degree in Computer Science and Engineering from MIT in 2017. She is the recipient of the Apple AI/ML Ph.D. fellowship, is a Rising Star in EECS and an R:SS and HRI Pioneer, has won best paper award at HRI 2020, and has interned at NVIDIA Research.

1 0

Aninda Manocha will present her FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on Monday, May 8th, 2023 at 3pm in CS 105
by Nicki Mahler 07 Apr '23

07 Apr '23

Aninda Manocha will present her FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on Monday, May 8th, 2023 at 3pm in CS 105. The members of her committee are as follows: Margaret Martonosi (Adviser); Readers: Kai Li and Luca Carloni (Columbia University); Examiners: David Wentzlaff and David August Everyone is invited to attend her talk. Title: "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" Abstract: Graph structures naturally and efficiently capture relationships between entities, such as individuals in a social network, pages in the World Wide Web, and amino acids in protein molecules. However, despite the advances in computing over the past few decades, graph processing algorithms continue to pose data supply challenges for current hardware. Modern networks are sparse and evergrowing, and therefore many graph applications are characterized by irregular memory access patterns that trouble data supply and memory management schemes. As a result, these applications are severely bottlenecked by several long-latency memory accesses to off-chip memory. My dissertation aims to tackle these bottlenecks with three main thrusts. First, this dissertation offers a hardware-software co-design that pairs automated compiler techniques to slice programs along bottleneck memory accesses with modest architectural modifications. This approach, GraphAttack, achieves a 2.9x speedup and 8.6x gain in energy efficiency over modern latency tolerance approaches. Second, this dissertation proposes tailored cache management policies supported with flexible hardware to improve memory bandwidth pressure and minimize off-chip memory accesses. This domain-specific memory hierarchy approach, Graphfire, yields a 1.3x speedup over state-of-the-art cache management policies and 63.3x speedup when scaling up to 64 cores. Third, this dissertation introduces architectural support for operating system (OS) techniques to efficiently identify data worth backing with huge pages and alleviate significant address translation overheads that occur in irregular applications. This results in 1.2-1.4x speedups over base pages alone when backing only 1-4% of the application footprint with huge pages. The end of Moore’s Law has pushed computer architects to design and develop hardware and software to accelerate compute and aim for aggressive power and performance targets. Despite these advances, data supply remains a challenge, particularly for memory-bound graph applications. Embracing hardware-software co-designs, my dissertation attacks the memory bottlenecks in graph processing workloads from multiple perspectives of the hardware-software stack to enable optimal performance. The techniques presented demonstrate the need to design properly equipped in-order technology and provide a path forward for tailored memory management for irregular applications in order to target a world where energy savings are critical for performance and scalability.

1 0

Decenter Seminar: Joachim Matthias Neu, Monday, March 27 at 3:30
by Emily C. Lawrence 07 Apr '23

07 Apr '23

Decenter Seminar Speaker: Joachim Matthias Neu, Stanford University Day: Monday, March 27, 2023 Time: 3:30-4:30pm Location: CS 105 Host: Pramod Viswanath Title: Internet-Scale Consensus In The Blockchain Era Abstract: Blockchains have ignited interest in Internet-scale consensus as a fundamental building block for decentralized applications and services, which in turn promise more egalitarian access and improved robustness to faults and abuse. While consensus has been studied in distributed systems for decades, Internet-scale consensus requires a fundamental rethinking of models, desiderata, and protocols. Participants are no longer a handful of computers in one company's data centers, but numerous mistrustful entities distributed across the Internet. I will discuss two examples of challenges and solutions for this new setting: (1) Ethereum, the second largest blockchain, aims to strengthen consensus liveness under open participation where parties come and go at will, and to strengthen safety to enable accountability in case of any safety violation. However, we show that no traditional single-ledger protocol can satisfy both strengthened properties simultaneously. To resolve this dilemma, we develop the multi-ledger consensus paradigm that is now the security design-specification for Ethereum. A by-product of this work are attacks on Ethereum's consensus protocol that prompted design changes. (2) Traditional network models do not capture rate constraints on communication and processing, leaving popular "provably secure" protocols vulnerable to attack. We show via a new queuing-based model how to schedule message handling securely. Our policies are simple enough to be forward-deployed at Internet service providers via a system that can also protect traffic of applications beyond blockchains. Bio: Joachim Neu is a PhD candidate at Stanford University with David Tse. His research focuses on the science and engineering of Internet-scale consensus as a fundamental building block for decentralized systems, using tools from distributed systems, applied probability, networking and communications, and applied cryptography. While a Masters student at Technical University of Munich, he worked in information and coding theory. Joachim has received the Protocol Labs PhD Fellowship and the Stanford Graduate Fellowship.

1 2

CS Colloquium Speaker Mae Milano TODAY at 12:30pm
by Emily C. Lawrence 06 Apr '23

06 Apr '23

Speaker: Mae Milano, University of California, Berkeley Date: Thursday, April 6 Time: 12:30pm EST Location: CS 105 Host: Andrew Appel Event page: https://www.cs.princeton.edu/events/26344 Title: Programming Distributed Systems Abstract: Our interconnected world is increasingly reliant on distributed systems of unprecedented scale, serving applications which must share state across the globe. And, despite decades of research, we're still not sure how to program them! In this talk, I'll show how to use ideas from programming languages to make programming at scale easier, without sacrificing performance, correctness, or expressive power in the process. We'll see how slight tweaks to modern imperative programming languages can provably eliminate common errors due to replica consistency or concurrency---with little to no programmer effort. We'll see how new language designs can unlock new systems designs, yielding both more comprehensible protocols and better performance. And we'll conclude by imagining together the role that a new cloud-centric programming language could play in the next generation of distributed programs. Bio: Mae Milano is a postdoctoral scholar at UC Berkeley working at the intersection of Programming Languages, Distributed Systems, and Databases. Her work has appeared at top-tier venues including PLDI, OOPSLA, POPL, VLDB, and TOCS, and has attracted the attention of the Swift language team. She is a recipient of the NDSEG Fellowship, has won several awards for her writing and service, and is a founding member of the Computing Connections Fellowship's selection committee (https://computingconnections.org/).

1 0

Yuting Yang will present her FPO "Exploiting Program Representation with Shader Applications" on Wednesday, May 3, 2023 at 3:00 PM in CS 402.
by CS Grad Department 05 Apr '23

05 Apr '23

Yuting Yang will present her FPO "Exploiting Program Representation with Shader Applications" on Wednesday, May 3, 2023 at 3:00 PM in CS 402. Location: CS 402 The members of Yuting’s committee are as follows: Examiners: Adam Finkelstein (Adviser), Szymon Rusinkiewicz, Jia Deng Readers: Felix Heide, Connelly Barnes (Adobe Research) A copy of her thesis will be available upon request, two weeks before the FPO. Please email [ mailto:gradinfo@cs.princeton.edu | gradinfo(a)cs.princeton.edu ] if you would like a copy of the thesis. Everyone is invited to attend her talk. Abstract follows below: Programs are widely used in content creation. For example, artists design shader programs to procedurally render scenes and textures, while musicians construct “synth” programs to generate electronic sound. While the generated content is typically the focus of attention, the programs themselves offer hidden potential for transformations that can support untapped applications. In this thesis, we will discuss four projects that exploit the program structure to automatically apply machine learning or math transformations as if they were manually designed by domain experts. First, we propose a compiler framework to automatically approximate the convolution of an arbitrary program with a Gaussian kernel in order to smooth the program for visual antialiasing. Second, we explore the benefit of program representation in deep-learning tasks by proposing to learn from program traces of procedural fragment shaders – programs that generate images. We next describe a compiler-based framework with novel math rules to extend reverse mode automatic differentiation so as to provide accurate gradients for arbitrary discontinuous programs. Finally, we extend the differentiation framework to audio “synth” programs so as to match the acoustic properties of a provided sound clip. In each of these settings, we demonstrate the benefit of exploiting the program structure to generalize hand-crafted techniques to arbitrary programs.

1 0

Princeton Robotics Seminar: Scott Kuindersma, Friday April 7
by Emily C. Lawrence 04 Apr '23

04 Apr '23

Speaker: Scott Kuindersma, Senior Director of Robotics Research, Boston Dynamics Date: Friday, April 7, 2023 Time: 11am ET Location: Computer Science 105 Event page: [ https://robo.princeton.edu/seminar | https://robo.princeton.edu/seminar ] Title: Taskable Agility: Making Useful Dynamic Behavior Easier to Create Abstract: In this talk, I will provide some insights and observations from our recent work on Atlas, the world's most dynamic humanoid robot. I'll cover the core technical ideas that have made an impact for us over the past few years and share my thoughts about the future for robots like Atlas. Bio: Scott Kuindersma is the Senior Director of Robotics Research at Boston Dynamics where he leads behavior research on Atlas. Prior to joining Boston Dynamics, he was an Assistant Professor of Engineering and Computer Science at Harvard. Scott’s research explores intersections of machine learning and model-based control to improve the capabilities of humanoids and other dynamic mobile manipulators.

1 0

ECE Seminar: Kuan Fang TODAY at 4:30pm
by Emily C. Lawrence 04 Apr '23

04 Apr '23

ECE SEMINAR Speaker: Kuan Fang, University of California, Berkeley Title: Scalable Robot Intelligence: Self-Supervised Learning through Generation Day: Tuesday, April 4, 2023 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: Chi Jin Abstract: Rapid advances in deep learning have resulted in promising techniques for robots to boost their capabilities to perceive, reason, and act by leveraging large models and massive datasets. However, the scalability of existing robot learning methods is severely limited by the manual labor and domain knowledge that humans can provide. To acquire general-purpose skills for solving a broad range of tasks, intelligent robots need scalable methods to collect and learn from rich data without extensive human supervision. In this talk, I will present my research on scaling up robot learning through the autonomous generation of environments, goals, and tasks. I will start by describing how to leverage procedural content generation for learning robust skills that can handle the variety and uncertainty of the real world. Then I will present algorithms that train robots to effectively reuse skills learned from prior experiences for novel sequential tasks by learning to generate reachable subgoals. Finally, I will demonstrate how to enable robots to discover a repertoire of novel skills by adaptively generating tasks during training. The acquired skills can be used for solving a variety of complex tasks such as tool use and sequential manipulation based on raw sensory inputs. Bio: Kuan Fang is a postdoctoral researcher in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, working with Sergey Levine. He received his Ph.D. degree in Electrical Engineering from Stanford University, advised by Fei-Fei Li and Silvio Savarese. His research interests lie at the intersection of robotics, computer vision, and machine learning, with a focus on developing data-driven methods to enable intelligent robots to operate in unstructured environments. He is a recipient of the Stanford Graduate Fellowship and the Computing Innovation Fellowship.

1 0