- talks - lists.cs.princeton.edu

Wei Hu will present his FPO "Understanding Deep Learning via Analyzing Dynamics of Gradient Descent" on Friday, July 23, 2021 at 9:30AM via zoom.
by jfarquer＠cs.princeton.edu 13 Jul '21

13 Jul '21

Wei Hu will present his FPO "Understanding Deep Learning via Analyzing Dynamics of Gradient Descent" on Friday, July 23, 2021 at 9:30AM via Zoom. Zoom link: https://princeton.zoom.us/j/2180408086 The members of Wei's committee are as follows: Sanjeev Arora (Adviser); Readers: Elad Hazan and Jason Lee; Examiners: Mark Braverman, Chi Jin and Sanjeev Arora. A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: The phenomenal successes of deep learning build upon the mysterious abilities of gradient-based optimization algorithms. Not only can these algorithms often successfully optimize complicated non-convex training objectives, but the solutions found can also generalize remarkably well to unseen test data despite significant over-parameterization of the models. Classical approaches in optimization and generalization theories that treat empirical risk minimization as a black box are insufficient to explain these mysteries in modern deep learning. This dissertation illustrates how we can make progress toward understanding optimization and generalization in deep learning by a more refined approach that opens the black box and analyzes the dynamics taken by the optimizer. In particular, we present several theoretical results that take into account the learning dynamics of the gradient descent algorithm. In the first part, we provide global convergence guarantees of gradient descent for training deep linear networks under various initialization schemes. Our results characterize the effect of width, depth and initialization on the speed of optimization. In addition, we identify an auto-balancing effect of gradient flow, which we prove to hold generally in homogeneous neural networks (including those with ReLU activation). In the second part, we study the implicit regularization induced by gradient descent, which is believed to be the key to mathematically understanding generalization in deep learning. We present results in both linear and non-linear neural networks, which characterize how gradient descent implicitly favors simple solutions. In the third part, we focus on the setting where neural networks are over-parameterized to have sufficiently large width. Through the connection to neural tangent kernels, we perform a fine-grained analysis of optimization and generalization, which explains several empirically observed phenomena. Built on these theoretical principles, we further design a new simple and effective method for training neural networks on noisily labeled data.

1 0

Raghuvansh Raj Saxena will present his FPO "Communication Complexity: For Coding, Commerce, and Currents" on Wednesday, July 21, 2021 at 12 noon via Zoom.
by jfarquer＠cs.princeton.edu 13 Jul '21

13 Jul '21

Raghuvansh Raj Saxena will present his FPO "Communication Complexity: For Coding, Commerce, and Currents" on Wednesday, July 21, 2021 at 12 noon via Zoom. Zoom Link: https://princeton.zoom.us/j/91504292936 The members of his Committee are as follows: Gillat Kol (Adviser); Readers: Matt Weinberg and Mark Braverman; Examiners: Ran Raz, Bernard Chazelle and Gillat Kol. A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is invited to attend the talk. Abstract follows below: Since its introduction in Yao's seminal paper [Yao79], communication complexity has revolutionized theoretical computer science. Indeed, not only has communication complexity been successful in areas where communication is an important resource, but it has also resulted in breakthroughs in areas that seemingly have nothing to do with communication! This is because bounds on communication can often be translated into bounds on other resources of interest, such as memory usage and circuit size. As the title suggests, we consider the role of communication complexity in the study of three different areas in theoretical computer science: Coding: Part I is about interactive error correcting codes, used to make interactive protocols resilient to noise. I study the noise tolerance of popular two-party and multi-party communication channels, constructing communication efficient codes with high noise tolerance for some of these channels, and proving that noise tolerance implies large communication overheads for others. Commerce: Part II studies communication problems in auction design, where the participating parties have their own interests and may not follow the protocol unless incentivized to do so. A common theme of this part is establishing tight separations between different classes of auctions, and understanding how various aspects of an auction affect its communication complexity. Currents: Part III covers my work on the memory requirement of graph streaming algorithms, which are algorithms whose input is a massive graph that they can only scan a few times. While algorithms that scan their input once are by now well understood, the situation is different for multi-pass streaming algorithms, which are the focus of this part of the thesis. Interestingly, each of these areas imposes additional requirements on the classical model of communication complexity: in interactive coding, we study communication complexity under noise; in auction design, we study communication complexity with incentives; and in multi-pass streaming, we consider communication complexity with few rounds of communication. While these three topics are seemingly very different, they all have strong connections with communication complexity, and studying them together often leads to deep insights.

1 0

Divyarthi Mohan will present her FPO "Simplicity and Optimality in Algorithmic Economics: Multi-Item Auctions and Social Learning" on Wednesday, July 14, 2021 at 2:30PM via Zoom.
by jfarquer＠cs.princeton.edu 07 Jul '21

07 Jul '21

Divyarthi Mohan will present her FPO "Simplicity and Optimality in Algorithmic Economics: Multi-Item Auctions and Social Learning" on Wednesday, July 14, 2021 at 2:30PM via Zoom. Zoom Link: https://princeton.zoom.us/j/94629442885?pwd=SURmMDBwYlg3YUp5YVVCMENvMEUrZz09 The members of Divya’s committee are as follows: Matt Weinberg ( Adviser), Readers: Mark Braverman, Nicole Immorlica (Microsoft Research); Examiners: Bernard Chazelle, Ran Raz, Matt Weinberg A copy of her thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is invited to attend her talk. Abstract follows below: Typically, algorithms are designed to find the best possible solution for a given input, while efficiently using resources. But in many settings, algorithms face an additional challenge: inputs are provided by strategic agents who may have different objectives than those of the algorithm designer. As a result, the field of algorithmic game theory emerged to understand and design algorithms in the presence of strategic agents. This thesis makes progress towards understanding simplicity in algorithmic game theory and when simplicity results in good outcomes; we focus on two main sub areas of algorithmic economics: multi-item auction theory and social learning. We consider the problem of revenue maximization when selling 𝑛 items to a single unit-demand buyer. Unfortunately, optimal auctions for selling multiple items can be un[1]reasonably complex and computationally intractable. Our work shows that simple mechanisms can achieve almost optimal revenue. We approach the tradeoffs of simplicity formally through the lens of computation and menu size. Our main result is a mechanism that gets a (1 − 𝜀)-approximation to the optimal revenue in time quasi-polynomial in 𝑛. We also study how information aggregates in networks through simple majority dynamics. We consider a model where agents make binary decisions, and each agent gets a private signal denoting which decision is better. Our main result proves that when the network is a tree formed according to the preferential attachment model, with high probability, the process stabilizes in a correct majority. Finally, we study a model of social learning between two Bayesian agents (a sender and a receiver) with limited/structured communication. In our model, to reflect real social exchanges, agents simply share a signal (or anecdote) they observed instead of sending arbitrary messages. A key result in our work shows that, when the agents have a fundamental difference, the optimal communication scheme can vary drastically (from sending the most representative anecdote to an extremely biased anecdote) depending on whether the receiver can observe the communication scheme used by the sender.

1 0

REVISED ABSTRACT: Nayana Prasad Nagendra will present her FPO "Improving Instruction Cache Performance For Modern Processors With Growing Workloads" on Wednesday, June 30, 2021 at 10AM via Zoom.
by jfarquer＠cs.princeton.edu 25 Jun '21

25 Jun '21

Nayana Prasad Nagendra will present her FPO "IMPROVING INSTRUCTION CACHE PERFORMANCE FOR MODERN PROCESSORS WITH GROWING WORKLOADS" on Wednesday, June 30, 2021 at 10AM via Zoom. Zoom link: https://princeton.zoom.us/j/92249108132 The members of her committee are as follows: David August (Adviser); Readers: David August and Tipp Moseley (Google); Examiners: David August, David Wentzlaff, Sharad Malik. A copy of her thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is invited to attend the talk. Abstract follows below: "The performance of warehouse-scale computing has a considerable financial and environmental impact. Our work with Google shows that more than two-thirds of the CPU cycles are wasted even after the extensive optimizations used today. Of these wasted cycles, nearly 20% are due to the processor front-end. In the first part of this talk, I will present the details of our fleet-wide study of Google workloads, a study we believe to be the first to explore front-end behavior at the warehouse scale. In the second part of this talk, I will present EMISSARY, a new instruction cache replacement algorithm designed to reduce these observed front-end stalls. By exploiting the fact that not all cache misses cause stalls, EMISSARY reserves cache capacity for lines whose misses cause the most performance impact. EMISSARY outperforms all measured instruction cache replacement algorithms in both performance and energy consumption. In some cases, it produces speedups that exceed Belady's OPT, the perfect-knowledge minimal miss ideal cache replacement algorithm."

1 0

Nayana Prasad Nagendra will present her FPO "IMPROVING INSTRUCTION CACHE PERFORMANCE FOR MODERN PROCESSORS WITH GROWING WORKLOADS" on Wednesday, June 30, 2021 at 10AM via Zoom.
by jfarquer＠cs.princeton.edu 25 Jun '21

25 Jun '21

Nayana Prasad Nagendra will present her FPO "IMPROVING INSTRUCTION CACHE PERFORMANCE FOR MODERN PROCESSORS WITH GROWING WORKLOADS" on Wednesday, June 30, 2021 at 10AM via Zoom. Zoom link: https://princeton.zoom.us/j/92249108132 The members of her committee are as follows: David August (Adviser); Readers: David August and Tipp Moseley (Google); Examiners: David August, David Wentzlaff, Sharad Malik. A copy of her thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is invited to attend the talk. Abstract follows below: The performance of warehouse-scale computing has a considerable financial and environmental impact. Our work with Google shows that more than two-thirds of the CPU cycles are wasted even after the extensive optimizations used today. Of these wasted cycles, nearly 20% are due to the processor front-end. In the first part of this talk, I will present the details of our fleet-wide study of Google workloads, a study we believe to be the first to explore front-end behavior at the warehouse scale. In the second part of this talk, I will present EMISSARY, a new instruction cache replacement algorithm designed to reduce these observed front-end stalls. By exploiting the fact that not all cache misses cause stalls, EMISSARY reserves cache capacity for lines whose misses cause the most performance impact. EMISSARY outperforms all known cache replacement algorithms in both performance and energy consumption. In some cases, it produces speedups that exceed Belady's OPT, the perfect-knowledge minimal miss ideal cache replacement algorithm.

1 0

Note Time Change: ECE Princeton Wireless Distinguished Seminar Series: Peter Vetter, Nokia Bell Labs - Monday, June 28th at 10:45am
by Emily C. Lawrence 24 Jun '21

24 Jun '21

Please note time change – 10:45am ECE Princeton Wireless Distinguished Seminar Series Speaker: Peter Vetter, Nokia Bell Labs Title: 6G – The Essential Infrastructure to Augment Human Potential at Scale Day: Monday, June 28, 2021 Time: 10:45am ET Host: Yasaman Ghasempour Please register here: [ https://princeton.zoom.us/meeting/register/tJ0kdeuvrDMqGtfvJXLX1p22lKKQm3x4… | https://princeton.zoom.us/meeting/register/tJ0kdeuvrDMqGtfvJXLX1p22lKKQm3x4… ] Abstract: 5G will increase human productivity thanks to the critical connectivity between man and machines or vehicles. The deployment of sensors and augmented intelligence closes the loop and increasingly automates operations, thereby freeing humans to focus on the tasks that they do well. A digital world that perfectly captures the past and current states of the physical world is emerging, however, the connectivity with the human biological and cognitive worlds still needs to be addressed. We anticipate that in the 6G era, there will be a need to deploy sensors and cognitive technologies at a much larger scale with stringent deterministic connectivity requirements. In the talk, we will share our vision how 6G era will be defined by the symbiosis of digital, physical, and biological worlds with the goal to augment human potential and wellbeing. We will then address the challenges for the 6G network, which will be the essential infrastructure for the integration of these future capabilities. Bio: Peter Vetter has recently been nominated President of Bell Labs Core Research at Nokia and is Bell Labs Fellow. He is leading an eminent global research organization with the mission to invent game changing innovations that define the future of networks, including network systems and security, optical systems, radio systems, and platforms & ASICs. In his previous role as Head of Access Research Lab, he and his teams have realized several world-first system demonstrations in access and successfully transferred industry leading concepts into product. He was also co-founder of an internal venture that produced the first FTTH product in Alcatel. He received the degree of Physics Engineer from Gent University (Belgium) in 1986 and a PhD with Prof. H. Pauwels in 1991. After a post-doctoral fellowship with Prof. T. Uchida at Tohoku University (Japan), he joined the research center of Alcatel (now Nokia) in Antwerp in 1993. Since 2009, he has worked at Bell Labs in Murray Hill, New Jersey, and has been on the senior leadership team of Bell Labs since 2013. He has authored over a hundred international papers and presented keynotes and tutorials at major technical industry events.

1 0

ECE Princeton Wireless Distinguished Seminar Series: Peter Vetter, Nokia Bell Labs - Monday, June 28
by Emily C. Lawrence 22 Jun '21

22 Jun '21

ECE Princeton Wireless Distinguished Seminar Series Speaker: Peter Vetter, Nokia Bell Labs Title: 6G – The Essential Infrastructure to Augment Human Potential at Scale Day: Monday, June 28, 2021 Time: 10:30am ET Host: Yasaman Ghasempour Please register here: [ https://princeton.zoom.us/meeting/register/tJ0kdeuvrDMqGtfvJXLX1p22lKKQm3x4… | https://princeton.zoom.us/meeting/register/tJ0kdeuvrDMqGtfvJXLX1p22lKKQm3x4… ] Abstract: 5G will increase human productivity thanks to the critical connectivity between man and machines or vehicles. The deployment of sensors and augmented intelligence closes the loop and increasingly automates operations, thereby freeing humans to focus on the tasks that they do well. A digital world that perfectly captures the past and current states of the physical world is emerging, however, the connectivity with the human biological and cognitive worlds still needs to be addressed. We anticipate that in the 6G era, there will be a need to deploy sensors and cognitive technologies at a much larger scale with stringent deterministic connectivity requirements. In the talk, we will share our vision how 6G era will be defined by the symbiosis of digital, physical, and biological worlds with the goal to augment human potential and wellbeing. We will then address the challenges for the 6G network, which will be the essential infrastructure for the integration of these future capabilities. Bio: Peter Vetter is Head of the A-Lab (Access research lab) at Nokia Bell Labs and Bell Labs Fellow. He is leading an eminent global team with the mission to invent game changing innovations that define the future of mobile and fixed access. Under his leadership, he and his teams have realized several world-first system demonstrations in access and successfully transferred industry leading concepts into product. He was also co-founder of an internal venture that produced the first FTTH product in Alcatel. He received the degree of Physics Engineer from Gent University (Belgium) in 1986 and a PhD with Prof. H. Pauwels in 1991. After a post-doctoral fellowship with Prof. T. Uchida at Tohoku University (Japan), he joined the research center of Alcatel (now Nokia) in Antwerp in 1993. Since 2009, he has worked at Bell Labs in Murray Hill, New Jersey, and has been on the senior leadership team of Bell Labs since 2013. He has authored over a hundred international papers and presented keynotes and tutorials at major technical industry events.

1 0

ECE Princeton Wireless Distinguished Seminar Series - Ozge Koymen, Qualcomm, TODAY at 1:30pm
by Emily C. Lawrence 17 Jun '21

17 Jun '21

Speaker: Ozge Koymen, Qualcomm Title: Qualcomm: Past, Present and Future of 5G Millimeter Wave Day: Thursday, June 17, 2021 Time: 1:30 pm ET Host: Yasaman Ghasempour (ECE) Event page: [ https://ece.princeton.edu/events/qualcomm-past-present-and-future-5g-millim… | https://ece.princeton.edu/events/qualcomm-past-present-and-future-5g-millim… ] Please register here: https://princeton.zoom.us/meeting/register/tJYvcuupqzIrGda2rjcGma_8FzK3QRUa… j Abstract: After more than a decade of advanced R&D and ecosystem trials, commercial 5G mmWave service is now available in more than 55 U.S. cities and 160 areas in Japan. Looking forward, we expect 5G mmWave to expand into new geographic regions across the globe, and new device types and tiers will emerge to take full advantage of mmWave’s virtually unlimited capacity. On the research front, Qualcomm continues to push the technology boundaries of mmWave for 5G/6G by bringing new capabilities and enhancements. Join this seminar to: * Review the key technical achievements and milestones at Qualcomm that enabled the commercialization of 5G mmWave systems. * See our vision for 5G mmWave and the new opportunities it poises to bring for the broader ecosystem. * Learn about the mmWave capabilities and enhancements coming in 3GPP Release -17 and beyond (e.g. Integrated Access and Backhaul, 60GHz and beyond, IIOT, etc.). * Track the latest update on the global commercial rollout of 5G mmWave networks and devices. Bio: Ozge Koymen is a Senior Director of Technology at Qualcomm Technologies, Inc. where he has been since 2006. He has led the 5G millimeter-wave program within Qualcomm R&D since early 2015, from early conceptual evaluation to commercial deployment. His previous areas as a technical contributor includes Wireless Backhaul, Small Cells, LTE-D, LTE and UMB. Prior to Qualcomm, he was a member of Flarion Technologies developing a pioneering OFDMA cellular system, Flash-OFDM, during 2003-2006. His earlier work experience includes full-time and consulting work for Impinj, Inc. (2000-2003) and TRW (1996-2000). He received the B.S. in Electrical and Computer Engineering from Carnegie Mellon University in 1996 and the M.S. and Ph.D. in Electrical Engineering from Stanford University in 1997 and 2003, respectively.

1 0

Alexander Upjohn Beatson will present his FPO "Learned surrogates and stochastic gradients for accelerating numerical modeling, simulation, and design" on Monday, June 21, 2921 at 12:00PM via Zoom.
by jfarquer＠cs.princeton.edu 14 Jun '21

14 Jun '21

Alexander Upjohn Beatson will present his FPO "Learned surrogates and stochastic gradients for accelerating numerical modeling, simulation, and design" on Monday, June 21,2921 at 12:00PM via Zoom. Zoom link: <https://www.google.com/url?q=https://princeton.zoom.us/j/99943194862&sa=D&s ource=calendar&usd=2&usg=AOvVaw0Ms6SOxf6CBLWGJW0mt45e> https://princeton.zoom.us/j/99943194862 The members of Alexander's committee are as follows: Ryan Adams (adviser); Readers: Szymon Rusinkiewicz, Sigrid Adriaenssens; Examiners: Ryan Adams, Elad Hazan, Mark Girolami (Cambridge University) A copy of Alex's thesis is available upon request. Please email gradinfo(a)cs.princeton.edu <mailto:gradinfo@cs.princeton.edu> if you would like a copy of the thesis. Everyone is entitled to attend his talk. Abstract follows below: Numerical methods, such as discretization-based methods for solving ODEs and PDEs, allow us to model and design complex devices, structures, and systems. However, this is often very costly in terms of both computation and the time of the expert who must specify the physical model, the discretization, the solver, etc. In this talk I will present a line of work using deep learning and stochastic gradient estimation to speed up solution and optimization of systems characterized by such numerical methods. First, I will present composable energy surrogates, in which neural surrogates are trained to model a potential energy in sub-components or sub-domains of a PDE, and then composed or stitched together to solve a larger system by minimizing the sum of potentials across components. This allows surrogate modeling without costly ground-truth simulation of the full system, as training data are generated by performing finite element analysis with individual components. We show that these surrogates can accelerate simulation of parametric meta-materials and produce accurate macroscopic behavior when composed. Next, I will discuss randomized telescoping gradient estimators, which provide unbiased gradient estimators for objectives which are the limit of a sequence of increasingly accurate, increasingly costly approximations. These estimators represent the limit as a telescoping sum and sample linear combinations of terms to provide cheap unbiased estimates. We discuss conditions which permit finite variance and computation, optimality of certain estimators within this class, and application to problems in numerical modeling and machine learning. Finally, I will discuss meta-learned implicit PDE solvers, which allow a new API for surrogate modeling. These models condition on a functional representation of a PDE and its domain by directly taking as input the PDE constraint and a method which returns samples in the domain and on the boundary. By avoiding having to fix a parametric basis, this allows fitting surrogate models for classes of PDEs with arbitrarily varying geometry and governing equations. It also does not require supervision from expensive ground-truth or FEA solutions. We apply Meta-PDE to a nonlinear Poisson problem, and show it learns to solve PDEs accurately and quickly across different boundary conditions, governing equations, and problem geometries. The resulting meta-model solves these PDEs much faster than FEA methods which achieve similar accuracy.

1 0

ECE Princeton Wireless Distinguished Seminar Series - Ozge Koymen, Qualcomm, Thursday, June 17 at 1:30pm
by Emily C. Lawrence 11 Jun '21

11 Jun '21

Speaker: Ozge Koymen, Qualcomm Title: Qualcomm: Past, Present and Future of 5G Millimeter Wave Day: Thursday, June 17, 2021 Time: 1:30 pm ET Host: Yasaman Ghasempour (ECE) Event page: [ https://ece.princeton.edu/events/qualcomm-past-present-and-future-5g-millim… | https://ece.princeton.edu/events/qualcomm-past-present-and-future-5g-millim… ] Please register here: [ https://princeton.zoom.us/meeting/register/tJYvcuupqzIrGda2rjcGma_8FzK3QRUa… | https://princeton.zoom.us/meeting/register/tJYvcuupqzIrGda2rjcGma_8FzK3QRUa… ] Abstract: After more than a decade of advanced R&D and ecosystem trials, commercial 5G mmWave service is now available in more than 55 U.S. cities and 160 areas in Japan. Looking forward, we expect 5G mmWave to expand into new geographic regions across the globe, and new device types and tiers will emerge to take full advantage of mmWave’s virtually unlimited capacity. On the research front, Qualcomm continues to push the technology boundaries of mmWave for 5G/6G by bringing new capabilities and enhancements. Join this seminar to: * Review the key technical achievements and milestones at Qualcomm that enabled the commercialization of 5G mmWave systems. * See our vision for 5G mmWave and the new opportunities it poises to bring for the broader ecosystem. * Learn about the mmWave capabilities and enhancements coming in 3GPP Release -17 and beyond (e.g. Integrated Access and Backhaul, 60GHz and beyond, IIOT, etc.). * Track the latest update on the global commercial rollout of 5G mmWave networks and devices. Bio: Ozge Koymen is a Senior Director of Technology at Qualcomm Technologies, Inc. where he has been since 2006. He has led the 5G millimeter-wave program within Qualcomm R&D since early 2015, from early conceptual evaluation to commercial deployment. His previous areas as a technical contributor includes Wireless Backhaul, Small Cells, LTE-D, LTE and UMB. Prior to Qualcomm, he was a member of Flarion Technologies developing a pioneering OFDMA cellular system, Flash-OFDM, during 2003-2006. His earlier work experience includes full-time and consulting work for Impinj, Inc. (2000-2003) and TRW (1996-2000). He received the B.S. in Electrical and Computer Engineering from Carnegie Mellon University in 1996 and the M.S. and Ph.D. in Electrical Engineering from Stanford University in 1997 and 2003, respectively.

1 0