- talks - lists.cs.princeton.edu

Aninda Manocha will present her Pre FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on May 12, 2022 at 2pm via Zoom
by Nicki Mahler 06 May '22

06 May '22

Aninda Manocha will present her Pre FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on May 12, 2022 at 2pm via Zoom. Zoom link: [ https://princeton.zoom.us/j/96476041500 | https://princeton.zoom.us/j/96476041500 ] Committee: Margaret Martonosi (adviser), Readers: Kai Li and Luca Carloni, Examiners: David August and David Wentzlaff All are welcome to attend. 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. Therefore, graph processing algorithms are commonly used in many important big-data applications. However, modern accelerator-oriented, heterogeneous system designs are inefficient for graph applications because they are characterized by irregular memory access patterns that confound sophisticated instruction scheduling and caching mechanisms as well as traditional prefetching and speculation techniques. As a result, these applications are severely bottlenecked by several long-latency memory accesses to off-chip DRAM and address translation overheads. My thesis aims to tackle these bottlenecks with three main thrusts spanning the hardware-software stack. First, I propose a hardware-software co-design composed of specialized access tracking buffers and precise compiler-driven program slicing techniques. This design enables significant memory-level parallelism and performance gains by issuing and overlapping multiple long-latency memory accesses to amortize their cost. This approach outperforms out-of-order cores, do-all parallelism, prefetching, and prior decoupling approaches, achieving a 2.87x speedup and 8.61x gain in energy efficiency across a range of graph applications. Second, I present domain-specific cache management techniques that leverage tailored fetch, insertion, and replacement policies to alleviate memory bandwidth pressure and minimize off-chip DRAM accesses. This work yields up to 1.79x (geomean 1.3x) speedup over state-of-the-art cache management policies. Lastly, I describe application-specific operating system (OS) techniques to efficiently manage page sizes and mitigate address translation overheads. These mechanisms offer 77.3-96.3% the performance of ideal huge page usage, while requiring only 0.58-2.92% of the huge page resources that modern operating systems, namely Linux, require. Overall, my work demonstrates the need for architects to revisit and equip simple hardware with specialized data supply technology to accelerate memory-bound applications when Dennard scaling no longer helps performance and energy savings are critical. Furthermore, as the end of Moore’s Law has pushed computer architects to employ hardware accelerators in order to reach aggressive power and performance targets for compute-bound workloads, domain-specific cache hierarchy approaches and OS techniques for memory-bound workloads are becoming increasingly necessary. Together, the three proposals in my dissertation work attack the memory bottlenecks in graph applications from different perspectives of the hardware-software stack to enable optimal performance.

1 0

CS Colloquium Speaker: Raymond Mooney, Friday, May 6 at 12:30pm
by Emily C. Lawrence 06 May '22

06 May '22

CS Colloquium Speaker Raymond Mooney, University of Texas at Austin Friday, May 6 - 12:30pm Computer Science - Room 105 Host: Danqi Chen Event website: [ https://www.cs.princeton.edu/events/26198 | https://www.cs.princeton.edu/events/26198 ] Title: Dialog with Robots: Perceptually Grounded Communication with Lifelong Learning Abstract: Developing robots that can accept instructions from and collaborate with human users is greatly enhanced by an ability to engage in natural language dialog. Unlike most other dialog scenarios, this requires grounding the semantic analysis of language in perception and action in the world. Although deep-learning has greatly enhanced methods for such grounded language understanding, it is difficult to ensure that the data used to train such models covers all of the concepts that a robot might encounter in practice. Therefore, we have developed methods that can continue to learn from dialog with users during ordinary use by acquiring additional targeted training data from the responses to intentionally designed clarification and active learning queries. These methods use reinforcement learning to automatically acquire dialog strategies that support both effective immediate task completion as well as learning that improves future performance. Using both experiments in simulation and with real robots, we have demonstrated that these methods exhibit life-long learning that improves long-term performance. Bio: Raymond J. Mooney is a Professor in the Department of Computer Science at the University of Texas at Austin. He received his Ph.D. in 1988 from the University of Illinois at Urbana/Champaign. He is an author of over 180 published research papers, primarily in the areas of machine learning and natural language processing. He was the President of the International Machine Learning Society from 2008-2011, program co-chair for AAAI 2006, general chair for HLT-EMNLP 2005, and co-chair for ICML 1990. He is a Fellow of AAAI, ACM, and ACL and the recipient of the Classic Paper award from AAAI-19 and best paper awards from AAAI-96, KDD-04, ICML-05 and ACL-07. This talk will be recorded and live-streamed at [ https://mediacentrallive.princeton.edu/ | https://mediacentrallive.princeton.edu/ ]

1 1

Zhijun Zhang will present his General Exam "Rounds vs Communication Tradeoffs for Maximal Independent Sets" on May 12, 2022 at 10am via Zoom
by Nicki Mahler 05 May '22

05 May '22

Zhijun Zhang will present his General Exam on May 12, 2022 at 10am via Zoom. Zoom link: [ https://princeton.zoom.us/my/zhijunz | https://princeton.zoom.us/my/zhijunz ] The members of his committee are as follows: Gillat Kol (adviser), Mark Braverman, and Ran Raz Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so. Abstract: I will survey my work in the fields of distributed sketching and interactive coding. Distributed sketching: The shared blackboard model is (arguably) the most basic model for distributed computing. It can also be viewed as a multi-round version of the well-studied sketching model. My recent work gives the first super-constant round lower bounds for some fundamental problems like MIS (minimal independent set), approximate matching, and approximate vertex cover, in the shared blackboard model. These bounds are tight (at least in some settings). Interactive coding: In the classical setting for error correcting codes, one party (the “sender”) wishes to send a message x to a second party (the “receiver”) over a noisy channel. We consider the coding question over the two-way noisy channel. Here, the sender and receiver can communicate back-and-forth with the goal of the receiver learning x. While it is well known that no binary error correcting code over the (one-way) bit flipping channel can tolerate more than 1/4 fraction of adversarial errors, my work shows that error resilience strictly better than 1/4 can be obtained over the two-way channel. Reading List: [ https://docs.google.com/document/d/1aFSXO95bm5tg7SikAOg2GpAC03Iq4uxgRmjWHN5… | https://docs.google.com/document/d/1aFSXO95bm5tg7SikAOg2GpAC03Iq4uxgRmjWHN5… ]

1 0

Yunyun Wang will present her General Exam "Controllable Speech Representation Learning via Voice Conversion" via Zoom
by Louis W. Riehl 05 May '22

05 May '22

Yunyun Wang will present her General Exam "Controllable Speech Representation Learning via Voice Conversion" on Thursday, May 12, 2022 at 3 pm via Zoom Committee Members: Adam Finkelstein (advisor), Karthik Narasimhan, Olga Russakovsky Abstract: Speech representation learning transforms speech into features that are suitable for downstream tasks, such as speech recognition, phoneme classification or speaker identification. For such recognition tasks, a representation can be lossy (non-invertible), which is typical of BERT-like, self-supervised models. However, when used for synthesis tasks, we find these lossy representations prove to be insufficient to plausibly reconstruct the input signal. This paper introduces a method for invertible and controllable speech representation learning based on disentanglement. The representation can be decoded into a signal perceptually identical to the original. Moreover, its disentangled components (speech content, pitch, speaker identity and energy) can be controlled independently to alter the synthesis result. Our model builds upon a zero-shot model called AutoVC-F0 trained for voice conversion – where the goal is to modify an audio recording containing the voice of one speaker so that the identity sounds like that of another speaker without altering the speech content. Building on AutoVC-F0, we introduce alteration invariant content loss (AIC loss) as well as adversarial training (GAN). Through objective measures and subjective tests, we show that our formulation offers significant improvement in voice conversion sound quality as well as more precise control over the disentangled features. Reading List: [ https://docs.google.com/document/d/1YTGrKke9HYHa7bxtNiyTyHpArrwSEXulPmrb_Dw… | https://docs.google.com/document/d/1YTGrKke9HYHa7bxtNiyTyHpArrwSEXulPmrb_Dw… ] 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

Jennifer Lam will present her General Exam "Thermopylae: Exploiting Skew in High Throughput Distributed Databases" on Wednesday, May 11, 2022 at at 2:00 PM in Friend 202 and via Zoom.
by Louis W. Riehl 05 May '22

05 May '22

Jennifer Lam will present her General Exam "Thermopylae: Exploiting Skew in High Throughput Distributed Databases" on Wednesday, May 11, 2022 at 2:00 PM in Friend 202 and via Zoom. Zoom link: [ https://princeton.zoom.us/j/95920790720 | https://princeton.zoom.us/j/95920790720 ] Committee Members: Wyatt Lloyd (advisor), Michael Freedman, Amit Levy Abstract: Distributed databases scale out by spreading data across many machines, which allows them to support large-scale applications whose data is too large to fit on a single-machine database. This scalability introduces a dilemma, however, as the throughput of distributed databases is much lower than that of modern single-machine databases for skewed workloads. This paper eliminates this dilemma as much as possible. We present Thermopylae, the first distributed database that is able to both support large-scale applications and match its throughput to single-machine databases under skewed workloads. Central to Thermopylae is a novel hybrid architecture that embeds a high-performance single-machine database into a highly scalable distributed database. Thermopylae applies a specialized concurrency control protocol designed for its hybrid architecture. Our evaluation shows that Thermopyale achieves orders of magnitude better performance than a state-of-the-art distributed database and closely matches its throughput to a modern single-machine database under skewed workloads. Reading List: [ https://docs.google.com/document/d/1JyKq72qi2xchJu8OH710RTF2m41E8Tf2hUgzCG-… | https://docs.google.com/document/d/1JyKq72qi2xchJu8OH710RTF2m41E8Tf2hUgzCG-… ] Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

UPDATE Sayash Kapoor will present his General Exam "Leakage and the reproducibility crisis in ML-based science" on Friday, May 6, 2022 at 9:00 AM VIA ZOOM ONLY
by Louis Riehl 05 May '22

05 May '22

Sayash Kapoor will present his General Exam "Leakage and the reproducibility crisis in ML-based science" on Friday, May 6, 2022 at 9:00 AM via zoom only. Zoom link: https://princeton.zoom.us/j/99164890280 Committee Members: Arvind Narayanan (advisor), Brandon Stewart, Olga Russakovsky Abstract: The use of Machine Learning (ML) methods for prediction and forecasting has become widespread across the quantitative sciences. However, there are many known methodological pitfalls, including data leakage, in ML-based science. In this work, we systematically investigate reproducibility issues in ML-based science. We show that data leakage is indeed a widespread problem and has led to severe reproducibility failures. Specifically, through a survey of literature in research communities that adopted ML methods, we find 15 fields where errors have been found, collectively affecting 304 papers and in some cases leading to wildly overoptimistic conclusions. Based on our survey, we present a fine-grained taxonomy of 8 types of leakage that range from textbook errors to open research problems. We argue for fundamental methodological changes to ML-based science so that cases of leakage can be caught before publication. To that end, we propose model cards for reporting scientific claims based on ML models and find that this would address all types of leakage identified in the 304 papers with errors in our survey. To investigate the impact of reproducibility errors and the efficacy of our model cards, we undertake a reproducibility study in a field where complex ML models are believed to vastly outperform older statistical models such as Logistic Regression (LR): Civil War prediction. We find that all papers claiming the superior performance of complex ML models compared to LR models fail to reproduce due to data leakage, and complex ML models don’t perform substantively better than decades-old LR models. While none of these errors could have been caught by reading the papers, our model cards enable the detection of leakage in each case. Reading List: https://docs.google.com/document/d/10Zt4UQTATXKb-aEOmNFi99xu4MbQiHcAXWc1hS3… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Jonathan Balkind will present his FPO "Open Source Platforms for Enabling Full-Stack Hardware-Software Research" on Monday, May 9, 2022 at 10am via Zoom
by Nicki Mahler 02 May '22

02 May '22

Jonathan Balkind will present his FPO "Open Source Platforms for Enabling Full-Stack Hardware-Software Research" on Monday, May 9, 2022 at 10am via Zoom. Zoom link: [ https://ucsb.zoom.us/j/83935399161?pwd=di9BRVRNMVJxNDAyZWUrSXE0V2VPZz09 | https://ucsb.zoom.us/j/83935399161?pwd=di9BRVRNMVJxNDAyZWUrSXE0V2VPZz09 ] His committee is as follows: David Wentzlaff (adviser), Readers: Amit Levy, Christopher Batten (Cornell), and David Wentzlaff; Examiners: Margaret Martonosi, Aarti Gupta, and David Wentzlaff All are welcome to attend. Abstract: Hardware-software research routinely starts with modelling. High-level models can capture much of the desired system behaviour needed to begin evaluating architectural research ideas. However, when those ideas come closer to realisation in real chips or they focus on cross-layer optimisations, higher-fidelity modelling tools must be brought to bear in evaluating them. This dissertation focuses on the development of two such tools, intended to enable research across the hardware-software stack. The tools described in this dissertation have been open-sourced in both their designs and characterisation in order to enable others to evaluate their research ideas with high fidelity. The first tool is OpenPiton, a full-stack manycore research platform and the worldâ€™s first open-source, general-purpose, multithreaded manycore processor. OpenPiton provides a toolbox for researchers to design new manycore processors incorporating their chosen research modifications. This toolbox includes scripts and tools necessary to produce a variety of processor configurations and take those instances all the way to realisation in real silicon chips. By embracing open-source design, the system supports modification at all levels of the stack, from underlying circuits to software running above, and as a result has been adopted in a variety of research areas across computing. The second tool presented in this dissertation is the Bring Your Own Core (BYOC) platform. BYOC is designed to enable heterogeneous-ISA systems research in architecture, systems software, and languages. BYOC brings together cores of multiple ISAs in a single, cache-coherent, manycore system and introduces a new interface for new cores to be rapidly connected to the system. This dissertation describes the integration of several new cores into OpenPiton to build this BYOC platform and the means taken to specifically support these cores existing together in a heterogeneous configuration, such that they can communicate via high-bandwidth and low-latency cache-coherent shared memory. This dissertation also details a characterisation of an FPGA prototype named Dicho, built with a combination of an OpenSPARC T1 SPARC V9 core and an Ariane RISC-V core.

1 0

Vivien Nguyen will present her General Exam "Line Drawing Simplification: Non-Photorealistic Rendering from 3D Models" on Monday, May 9, 2022 at 2:00 PM in CS 401 and via zoom.
by Louis W. Riehl 02 May '22

02 May '22

Vivien Nguyen will present her General Exam "Line Drawing Simplification: Non-Photorealistic Rendering from 3D Models" on Monday, May 9, 2022 at 2:00 PM in CS 401 and via zoom. Zoom link: [ https://princeton.zoom.us/j/6409315927 | https://princeton.zoom.us/j/6409315927 ] Committee Members: Szymon Rusinkiewicz (advisor), Adam Finkelstein, Felix Heide. Abstract: When artists make line drawings corresponding to real-world objects, they may selectively choose to include, omit, merge, or modify certain lines depending on their stylistic or communicative goals. In this talk, I will discuss my work on simplifying line drawings resulting from non-photorealistic rendering (NPR) of 3D models. In the traditional NPR pipeline, a set of curves and their visibility are extracted, then rendered as a set of pixel-space lines. When the input mesh is polyhedral (i.e. non-smooth) or due to numerical instability, the curves extracted may be overly complex or present other artifacts. Even when the curves are correct, we may wish to remove or simplify them for performance or artistic reasons. I will first discuss prior work in both creating and simplifying line drawings. I will then describe how we use recent techniques in differentiable vector graphics for a learning-based approach to line drawing simplification, as well as the challenges in doing so. Reading List: [ https://docs.google.com/document/d/19qCcNJuXzFfkQY3n2XfhtEYYsS4u92PVnS51eZ0… | https://docs.google.com/document/d/19qCcNJuXzFfkQY3n2XfhtEYYsS4u92PVnS51eZ0… ] Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Minsung Kim will present his Pre FPO "Quantum and Quantum-Inspired Computation for Wireless Networks" on Friday, May 13, 2022 at 10am via Zoom
by Nicki Mahler 29 Apr '22

29 Apr '22

Minsung Kim will present his Pre FPO "Quantum and Quantum-Inspired Computation for Wireless Networks" on Friday, May 13, 2022 at 10am via Zoom. Zoom link: https://princeton.zoom.us/j/97528898987 The members of his committee are as follows: Examiners: Kyle Jamieson (adviser), Jennifer Rexford, and Yasaman Ghasempour (ECE); Readers: Ravi Netravali, Lin Zhong, and Davide Venturelli (NASA/USRA RIACS) All are welcome to attend. ============================================================== Abstract: A central design challenge for future generations of wireless networks is to meet the ever-increasing demand for wireless capacity. While significant progress has been made in designing advanced wireless technologies, the current computational capacity at base stations to support them has been consistently identified as the bottleneck, due to limitations in processing time. Quantum computing is a potential tool to overcome the tradeoff between the wireless performance and computational complexity. It exploits unique information processing capabilities based on quantum mechanics to perform fast calculations that are intractable by traditional digital methods. In this talk, I will present design directions of quantum compute-enabled base station systems in wireless networks and introduce our prototype systems that are implemented on real-world quantum processors. The prototypes are designed for quantum-accelerated near-optimal wireless signal processing in Multiple-Input Multiple-Output (MIMO) systems that could drastically increase wireless capacity for the cellular 5G New Radio roadmap, as well as in next generation wireless local area networks. I will provide design guidance of quantum, quantum-inspired classical, and classical-quantum hybrid optimization in the systems with underlying principles and technical details, and discuss future research directions based on the current challenges and opportunities.

1 0

Carlos Jimenez will present his General Exam "Consistency And Robustness Testing for Visual Question Answering" on Friday, May 6, 2022 at 2:30 PM in CS 401 and via zoom.
by Louis W. Riehl 29 Apr '22

29 Apr '22

Carlos Jimenez will present his General Exam "Consistency And Robustness Testing for Visual Question Answering" on Friday, May 6, 2022 at 2:30 PM in CS 401 and via zoom. Zoom link: [ https://princeton.zoom.us/j/96002443578 | https://princeton.zoom.us/j/96002443578 ] Committee Members: Karthik Narasimhan (advisor), Danqi Chen, Olga Russakovsky Abstract: We introduce CARETS as a new systematic test suite to measure consistency and robustness of modern visual question answering (VQA) models through a series of fine-grained capability tests. In contrast to existing VQA test sets, CARETS features balanced question generation to create pairs of instances to test models, with each pair focusing on a specific capability such as rephrasing, logical symmetry or image obfuscation. We evaluate six modern VQA systems on CARETS and identify several actionable weaknesses in model comprehension, especially with concepts such as negation, disjunction, or hypernym invariance. We show that even the most sophisticated models are sensitive to aspects such as swapping the order of terms in a conjunction or varying the number of answer choices mentioned in a question; underscoring the need for improved robustness and reliability of state-of-the-art machine learning models. Reading List: [ https://docs.google.com/document/d/1rnOlJVkDNgI7CJsjVN78sEPN-OIa3M9i661f_ol… | https://docs.google.com/document/d/1rnOlJVkDNgI7CJsjVN78sEPN-OIa3M9i661f_ol… ] 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