- talks - lists.cs.princeton.edu

Speaker-Li Erran Li, Tomorrow, Sept 5
by Emily Lawrence 04 Sep '18

04 Sep '18

Title: Machine Learning for Autonomous Driving: Recent Advances and Future Challenges Li Erran Li, Pony AI Wed Sept 5 12-1pm; Room 302 in CS building (Pizza served before the talk, and the speaker expressed interest in talking to grads) Abstract: Tremendous progresses have been made in applying machine learning to autonomous driving. However, there are fundamental challenges ahead. In this talk, I will present recent advances in applying machine learning to solving the perception, prediction, planning and control problems of autonomous driving. I will discuss key research challenges in learning more robust and abstract representations, scene understanding, behavior prediction, and decision-making in complex real-world scenarios. Bio: Dr. Li Erran Li is the chief scientist at Pony.ai and an adjunct professor at Columbia University. Most recently, he was with the perception team at Uber ATG. Before that, he was at Bell Labs. Dr. Li’s current research interests are machine learning, computer vision and learning-based robotics. Dr. Li has a PhD from the Computer Science Department at Cornell University. Dr. Li is an IEEE Fellow and an ACM Fellow.

1 0

Muhammad Shahbaz will present his FPO, " Enabling Programmable Infrastructure for Multi-tenant Data Centers " on Friday, 9/7/2018 at 11am in CS 302.
by Nicki Gotsis 31 Aug '18

31 Aug '18

Muhammad Shahbaz will present his FPO, " Enabling Programmable Infrastructure for Multi-tenant Data Centers " on Friday, 9/7/2018 at 11am in CS 302. The members of his committee are as follows: Adviser: Nick Feamster; Readers: Jennifer Rexford and Ben Pfaff (VMware Inc.); Examiners: Wyatt Lloyd, Michael Freedman, and Nick Feamster. Everyone is welcome to attend. A copy of his thesis is available in CS 310. Abstract follows below. Today’s data centers are large, shared infrastructures hosting hundreds of thousands of tenants over a vast network of servers. Operating these data centers frequently entails incorporating new infrastructure services that require customizing the behavior of its switches and exploiting unique characteristics of data centers to scale. Emerging programmable switch ASICs allow network operators to customize the behavior of physical switches. Yet, virtual switches running on servers in multitenant data centers are still fixed function and composed of large, complex software codebases. Modifying these switches requires both intimate knowledge of the switch codebase and extensive expertise in network protocol design, raising the bar for customizing these switches prohibitively high. In this dissertation, we address these challenges by, first, presenting the design and implementation of PISCES: a programmable, protocol-independent software switch derived from Open vSwitch (OVS), a fixed-function hypervisor switch, whose behavior is customized using P4. PISCES is not tethered to specific protocols; this independence makes it easy to add new features. We also show how the compiler can analyze the high-level specification to optimize forwarding performance. Our evaluation shows that PISCES performs comparably to OVS and that P4 programs for PISCES are about 40 times shorter than equivalent changes to OVS source code. Next, we demonstrate how such programmable switches help build scalable infrastructure services by exploiting the unique characteristics of data-center networks. We use these switches to address the multicast scalability problem in multi-tenant data centers and present the design of an infrastructure service, Elmo, that scales multicast by taking advantage of the data-center characteristics; specifically, the symmetric topology and short paths in a data center. In Elmo, a PISCES switch encodes multicast group information inside packets themselves, reducing the need to store the same information in hardware switches, which instead read the encoded information to route packets to recipients. In a three-tier data-center topology with 27,000 hosts, Elmo supports a million multicast groups using a 325-byte packet header, requiring as few as 1,100 multicast group-table entries on average in hardware switches, with a traffic overhead as low as 5% over ideal multicast.

1 0

Arpit Gupta will present his FPO "Flexible and Scalable Systems for Network Management" on Friday, 8/31/2018 at 10 am in CS 302.
by Nicki Gotsis 28 Aug '18

28 Aug '18

Arpit Gupta will present his FPO "Flexible and Scalable Systems for Network Management" on Friday, 8/31/2018 at 10 am in CS 302. The members of his committee are as follows: Adviser: Nick Feamster; Readers: Nick Feamster, Jennifer Rexford, and Walter Willinger (NIKSUN Inc.); Examiners: Marshini Chetty, Kyle Jamieson, and Nick Feamster. Everyone is invited to attend. A copy of his thesis is available in CS 310. Thesis abstract follows below. Our daily lives are heavily reliant upon Internet-connected devices, services, and applications. This reliance makes it more critical than ever that the underlying networks they depend on be reliable, performant, and secure. At the same time, the increasing complexity and diversity of today's devices, services, and applications have made network management tasks more complicated than ever. Modern network management mandates that operators can systematically monitor what is going on in their networks (network monitoring) and use this information to take real-time preventive or corrective actions (network control). Achieving these goals while also adhering to the limited compute and storage resources available on modern network devices poses signi\u000Ccant challenges. The contribution of this dissertation is the design and implementation of two systems that enable exible and scalable network monitoring and control. The networkmonitoring system, Sonata, collects and analyzes network tra c to infer various network events in real time. The network-control system, SDX, enables \u000Cne-grained reactive control actions for interdomain tra c without disrupting the existing routing protocols. For each of these two systems, the dissertation focuses on (i) the ab- stractions that allow network operators to express exible programs for both network monitoring and control; (ii) the algorithms that make the best use of limited compute and storage resources; and (iii) the systems that combine the high-level abstractions and the low-level algorithms and can be deployed in production settings. The lessons learned from this dissertation can help us design next-generation network-management systems. More concretely, unlike existing systems that rely solely on a single device-type, this dissertation shows that designing systems that can pool resources from a heterogeneous set of devices (targets) is critical for building exible and scalable network-management systems. It also demonstrates that as the networking technologies and protocols evolve rapidly with time, it is imperative to design modular systems that can swiftly catch up with these changes. Finally, this research also illustrates that it is crucial to select strategic locations (e.g., Internet exchange points) for deployment to drive innovations in Internet-wide tra c monitoring and control.

1 0

Pawel Przytycki will present his FPO "Algorithms for deciphering cancer genomes: from differential mutation to differential allele specific expression" on Friday, 8/31/2018 at 11:30 in CS 402.
by Nicki Gotsis 24 Aug '18

24 Aug '18

Pawel Przytycki will present his FPO "Algorithms for deciphering cancer genomes: from differential mutation to differential allele specific expression" on Friday, 8/31/2018 at 11:30 in CS 402. The members of his committee are as follows: Mona Singh (adviser); Joshua Akey (LSI) and Olga Troyanskaya (readers); Mona Singh, Ben Raphael, and Yibin Kang (MOL) (examiners). All are welcome to attend. A copy of his thesis is available in CS 310. The thesis abstract follows below. Large-scale cancer genome sequencing consortia have provided a huge influx of somatic mutation data across large cohorts of patients. Understanding how these observed genetic alterations give rise to specific cancer phenotypes represents a major aim of cancer genomics. In this dissertation, I present two methods for utilizing natural variation as a background for interpreting cancer genomes. In Chapter 2, I introduce di↵erential mutation analysis, a framework for uncovering cancer genes that compares the mutational profiles of genes across cancer genomes with natural germline variation across healthy individuals. I hypothesize that if a gene is less constrained with respect to variation across the healthy population, it may also be able to tolerate a greater amount of somatic mutation without experiencing a drastic detrimental functional change. I develop a fast and simple approach that uncovers genes that are di↵erentially mutated between cancer genomes and the genomes of healthy individuals. I demonstrate that my di↵erential mutation approach outperforms considerably more sophisticated approaches for discovering cancer genes. In Chapter 3, I propose the concept of di↵erential allele-specific expression to identify genes within an individual’s cancer whose allele-specific expression (ASE) differs from what is found in matched normal tissue, with the overall goal of uncovering genes whose regulation is altered via functional noncoding somatic mutations. I reason that since specific noncoding mutations usually occur on only one chromosome, they are expected to a↵ect only the expression of the allele derived from that chromosome. Thus, ASE is a potential avenue towards detecting cis mutations that lead to regulatory changes. I present three methods to identify di↵erential ASE in paired tumor-normal samples, and apply them to breast cancer tumor samples. I demonstrate that di↵erential ASE can detect dysregulation caused by nonsense mediated decay and copy number variation, that known cancer-related genes are enriched for di↵erential ASE, and that genes with cis noncoding mutations are enriched for difiii ferential ASE. Finally, I show that noncoding mutations in cis with genes exhibiting di↵erential ASE often disrupt known regulatory mechanisms. I thus conclude that di↵erential ASE is a powerful means for characterizing gene dysregulation due to cis noncoding mutations.

1 0

Arpit Gupta will present his FPO "Flexible and Scalable Systems for Network Management" on Friday, 8/31/2018 at 10 am in CS 302.
by Nicki Gotsis 23 Aug '18

23 Aug '18

Arpit Gupta will present his FPO "Flexible and Scalable Systems for Network Management" on Friday, 8/31/2018 at 10 am in CS 302. The members of his committee are as follows: Adviser: Nick Feamster; Readers: Nick Feamster, Jennifer Rexford, and Walter Willinger (NIKSUN Inc.); Examiners: Marshini Shetty, Kyle Jamieson, and Nick Feamster. Everyone is invited to attend. A copy of his thesis is available in CS 310. Thesis abstract follows below. Our daily lives are heavily reliant upon Internet-connected devices, services, and applications. This reliance makes it more critical than ever that the underlying networks they depend on be reliable, performant, and secure. At the same time, the increasing complexity and diversity of today's devices, services, and applications have made network management tasks more complicated than ever. Modern network management mandates that operators can systematically monitor what is going on in their networks (network monitoring) and use this information to take real-time preventive or corrective actions (network control). Achieving these goals while also adhering to the limited compute and storage resources available on modern network devices poses signi cant challenges. The contribution of this dissertation is the design and implementation of two systems that enable exible and scalable network monitoring and control. The networkmonitoring system, Sonata, collects and analyzes network trac to infer various network events in real time. The network-control system, SDX, enables ne-grained reactive control actions for interdomain trac without disrupting the existing routing protocols. For each of these two systems, the dissertation focuses on (i) the ab- stractions that allow network operators to express exible programs for both network monitoring and control; (ii) the algorithms that make the best use of limited compute and storage resources; and (iii) the systems that combine the high-level abstractions and the low-level algorithms and can be deployed in production settings. The lessons learned from this dissertation can help us design next-generation network-management systems. More concretely, unlike existing systems that rely solely on a single device-type, this dissertation shows that designing systems that can pool resources from a heterogeneous set of devices (targets) is critical for building exible and scalable network-management systems. It also demonstrates that as the networking technologies and protocols evolve rapidly with time, it is imperative to design modular systems that can swiftly catch up with these changes. Finally, this research also illustrates that it is crucial to select strategic locations (e.g., Internet exchange points) for deployment to drive innovations in Internet-wide trac monitoring and control.

1 0

Steven Goldfeder will present his FPO, "Off-chain protocols for cryptocurrencies" on Wednesday, 8/29/2018 at 2:00 PM in CS 402.
by Nicki Gotsis 23 Aug '18

23 Aug '18

Steven Goldfeder will present his FPO, "Off-chain protocols for cryptocurrencies" on Wednesday, 8/29/2018 at 2:00 PM in CS 402. The members of his committee are as follows: Adviser: Arvind Narayanan; Readers: Edward Felten and Rosario Gennaro (City College - CUNY); Examiners: Matt Weinberg, Mark Zhandry, and Arvind Narayanan Everyone is invited to attend. A copy of his thesis is available in CS 310. The thesis abstract follows below. The limits of Bitcoin’s scripting language motivate the need for o↵-blockchain protocols that extend the functionality of Bitcoin scripts. These protocols are run out-ofband by the transacting parties, but they are constructed in a manner that cryptographically binds them to on-chain scripts. Even with the advent of Ethereum, which provides a much richer scripting language, there are still privacy and scalability benefits to running o↵-chain protocols, even when on-chain analogs exist. We present o↵-chain protocols for a variety of applications as well as a general framework for o↵-chain smart contracts. An important tool for constructing o↵-chain protocols is threshold-signatures, a primitive that enables distributing the signing power of a given public key into n shares, such that at least k shares are required to produce a signature from that key. We show how to construct threshold signatures that are compatible with Bitcoin, and we then use this primitive to build o↵-chain protocols for privacy-preserving access control and escrow services. Turning to more advanced smart contracts, we present o↵-chain protocols for the fair-exchange of digital goods and services for payment. Finally, we present Arbitrum, a private and scalable smart contract system which enables running arbitrary smart contracts for which the code is executed o↵-chain and disputes are resolved on-chain.

1 0

Max Homilius will present his FPO, "Network-Based Prioritization of Disease Genes, Animal Models, and Drug Targets" on Friday, 8/24/2018 at 10am, 280 Lewis-Sigler Institute.
by Nicki Gotsis 20 Aug '18

20 Aug '18

Max Homilius will present his FPO, "Network-Based Prioritization of Disease Genes, Animal Models, and Drug Targets" on Friday, 8/24/2018 at 10am, 280 Lewis-Sigler Institute. The members of his committee are as follows: Readers: Barbara Engelhardt and Tilo Grosser (UPenn); Examiners: Mona Singh, Coleen Murphy (MOL), and Olga Troyanskaya (adviser). A copy of his thesis is available upon request. Everyone is invited to attend his talk. The talk abstract follows below. In living organisms, biomolecules interact in complex molecular networks that underlie cell function and whose dysregulation leads to disease. These networks thus provide a key lens for understanding the molecular basis of human disease as well as treatment development. In this thesis, I develop three network-based computational approaches for human disease research and gaining insight into the mode of action of drugs. At their core, these methods employ functional interaction networks, which provide a genome-wide view of biochemical and pathway-level interactions and summarize essential functional information derived from diverse and heterogeneous functional genomics experiments. First, I propose a network-based method that can detect critical genes and pathways targeted by a drug treatment from gene expression data even in the absence of large-scale expression di↵erences. This approach enables the analysis of low-dose drug screens, ranking potential targets and drug-perturbed biological processes with higher accuracy than prior network-based methods or gene-expression data alone. Furthermore, I present a method that by inferring and comparing genome-wide profiles for human diseases and animal model phenotypes identifies analogous disease models with high accuracy and more robustly than prior methods relying on shared gene content. This method allows to aggregate the wealth of existing model organism knowledge across multiple species and to identify related phenotypes and novel homologous genes of human diseases for experimental follow-up. Lastly, by constructing a joint tissue-specific classifier for human disease genes, we can significantly improve the prediction of associated genes for rare human diseases. This neural network-based approach makes use of a functional network embedding leveraging tissue-specific expression data and model organism phenotype information in a multi-label classification setting. Overall, the methods I developed provide data-driven, molecular-level solutions to major biological challenges relevant to human health.

1 0

Vicky Yao will present her FPO "Integrative network-based approaches to analyze genomics data" on Friday, 8/24/2018 at 1:00pm in 200 CIL.
by Nicki Gotsis 20 Aug '18

20 Aug '18

Vicky Yao will present her FPO "Integrative network-based approaches to analyze genomics data" on Friday, 8/24/2018 at 1:00pm in 200 CIL. The members of her committee are as follows: Readers: Mona Singh and Barbara Engelhardt; Examiners: Ryan Adams, Coleen Murphy (MOL), and Olga Troyanskaya (adviser). A copy of her thesis is available upon request. Everyone is invited to attend her talk. The talk abstract follows below. The generation of diverse genome-scale data across organisms and experimental conditions is becoming increasingly commonplace, creating unprecedented opportunities for understanding the molecular underpinnings of human disease. However, these large data are often noisy, highly heterogeneous, and lack the resolution required to study key aspects of metazoan complexity, such as tissue and cell-type specificity. Furthermore, targeted data collection and experimental verification is often infeasible in humans, underscoring the need for methods that can integrate -omics data, computational predictions, and biological knowledge across organisms. In this dissertation, I describe several novel, integrative computational approaches to address these challenges. First, I will describe a statistical and machine learning approach that takes advantage of high-quality neuron-specific molecular profiles of cells that vary in vulnerability to Alzheimer’s disease obtained in mouse to generate neuron-specific functional networks in human. We then combine these network models with human quantitative genetics data to prioritize likely Alzheimer’s disease candidates. Next, I present an in-depth analysis of all major adult C. elegans tissues and genome-wide expression predictions across 76 tissues and cell types. The tissue expression prediction method is one of the building blocks of diseaseQUEST, an integrative computational-experimental framework that combines human quantitative genetics with in silico functional network representations of model organism biology to systematically identify disease gene candidates. This framework leverages a novel semi-supervised Bayesian network integration approach to predict tissue- and cell-type specific functional relationships between genes in model organisms. We use diseaseQUEST to construct 203 tissue- and cell-type specific functional networks and predict candidate genes for 25 different human diseases and traits using C. elegans as a model system, with a particular focus on Parkinson’s disease. Finally, I will present a network-based approach that systematically identifies differential isoforminteractions. We apply this approach to the study of tissue and environment dynamics in Alzheimer’s disease. Together, these approaches provide a framework for addressing the challenges of data heterogeneity, noise, and biological resolution in human molecular data to better understand the etiology of human disease.

1 0

Friday: EE Seminar-11:30am, E-Quad B205- Andrea Goldsmith, Stanford University
by Emily Lawrence 14 Aug '18

14 Aug '18

Department of Electrical Engineering Seminar Speaker: Dr. Andrea Goldsmith Stanford University Title: Can machine learning trump theory in communication system design? Date: Friday, August 17, 2018 Time: 11:30AM Room: E-Quad, B205 Host: Prof. H. Vincent Poor & Prof. Yuxin Chen Abstract: Design and analysis of communication systems have traditionally relied on mathematical and statistical channel models that describe how a signal is corrupted during transmission. In particular, communication techniques such as modulation, coding and detection that mitigate performance degradation due to channel impairments are based on such channel models and, in some cases, instantaneous channel state information about the model. However, there are propagation environments where this approach does not work well because the underlying physical channel is too complicated, poorly understood, or rapidly time-varying. In these scenarios we propose a completely new approach to communication system design based on machine learning (ML). In this approach, the design of a particular component of the communication system (e.g. the coding strategy or the detection algorithm) utilizes tools from ML to learn and refine the design directly from training data. The training data that is used in this ML approach can be generated through models, simulations, or field measurements. We present results for three communication design problems where the ML approach results in better performance than current state-of-the-art techniques: signal detection without accurate channel state information, signal detection without a mathematical channel model, and joint source-channel coding of text. Broader application of ML to communication system design in general and to millimeter wave and molecular communication systems in particular is also discussed. ------------------------------------------------------------- Bio: Andrea Goldsmith is the Stephen Harris professor in the School of Engineering and a professor of Electrical Engineering at Stanford University. She also serves on Stanford's Presidential Advisory Board, University Budget Group, and Faculty Senate. She previously served as Chair of Stanford's Faculty Senate and as a member of Stanford's Commission on Graduate Education, Commission on Undergraduate Education, Committee on Research, Planning and Policy Board, and Task Force on Women and Leadership. She co-founded and served as Chief Technical Officer of Plume WiFi (formerly Accelera, Inc.) and of Quantenna (QTNA), Inc. She has also held industry positions at Maxim Technologies, Memorylink Corporation, and AT&T Bell Laboratories, and she currently chairs the Technical Advisory Boards of Interdigital Corp., Quantenna Communications, Cohere Communications, and Sequans. In the IEEE Dr. Goldsmith served on the Board of Governors for both the Information Theory and Communications societies. She has also been a Distinguished Lecturer for both societies, served as President of the IEEE Information Theory Society in 2009, founded and chaired the student committee of the IEEE Information Theory society, and chaired the Emerging Technology Committee of the IEEE Communications Society. She currently chairs the IEEE TAB committee on diversity and inclusion, and the Women in Technology Leadership Roundtable working group on metrics. Dr. Goldsmith is a member of the National Academy of Engineering and the American Academy of Arts and Sciences, a Fellow of the IEEE and of Stanford, and has received several awards for her work, including the IEEE ComSoc Edwin H. Armstrong Achievement Award as well as Technical Achievement Awards in Communications Theory and in Wireless Communications, the National Academy of Engineering Gilbreth Lecture Award, the IEEE ComSoc and Information Theory Society Joint Paper Award, the IEEE ComSoc Best Tutorial Paper Award, the Alfred P. Sloan Fellowship, the WICE Technical Achievement Award, and the Silicon Valley/San Jose Business Journal's Women of Influence Award. She is author of the book ``Wireless Communications'' and co-author of the books ``MIMO Wireless Communications'' and "Principles of Cognitive Radio," all published by Cambridge University Press, as well as an inventor on 28 patents. Her research interests are in information theory and communication theory, and their application to wireless communications and related fields. She received the B.S., M.S. and Ph.D. degrees in Electrical Engineering from U.C. Berkeley.

1 0

Jonathan Schneider will present his FPO "Learning Algorithms in Strategic Environments" on Wednesday, 8/15/2018 at 1pm in CS 402.
by Nicki Gotsis 09 Aug '18

09 Aug '18

Jonathan Schneider will present his FPO "Learning Algorithms in Strategic Environments" on Wednesday, 8/15/2018 at 1pm in CS 402. The members of his committee are as follows: Adviser: Mark Braverman; Readers: Matt Weinberg and Omri Weinstein (Columbia); Examiners: Bernard Chazelle, Elad Hazan, and Mark Braverman. A copy of his thesis is available upon request. Everyone is invited to attend his talk. The talk abstract follows below. Learning algorithms are often analyzed under the assumption their inputs are drawn from stochastic or adversarial sources. Increasingly, these algorithms are being applied in strategic settings, where we can hope for stronger guarantees. This thesis aims to understand the performance of existing learning algorithms in these settings, and to design new algorithms that perform well in these settings. This thesis is divided into three parts. In Part I, we address the question of how agents should learn to bid in repeated non-truthful auctions { and conversely, how should we design auctions whose participants are learning agents. In Part II, we study the dynamic pricing problem: the question of how should a large retailer learn how to set prices for a sequence of disparate goods over time, based on observing demands for goods at various prices. Previous work has demonstrated how to obtain O(log T) regret for this problem. We show how to achieve regret O(log log T), which is tight. Our algorithm uses ideas from integral geometry (most notably the concept of intrinsic volumes). Finally, in Part III, we study how to learn the ranking of a set of N items from pairwise comparisons that may be strategic or noisy. In particular, we design mechanisms for a variety of settings (choosing the winner of a round-robin tournament, aggregating the top-K items under the strong stochastic transitivity noise model) which outperform the naive rule of ranking items according to the total number of pairwise comparisons won.

1 0