Quick links

CS Department Colloquium Series

Resource-Efficient Execution for Deep Learning

Date and Time
Wednesday, March 10, 2021 - 4:30pm to 5:30pm
Location
Zoom Webinar (off campus)
Type
CS Department Colloquium Series
Host
Kyle Jamieson

Please register here.


Deepak Narayanan
Deep Learning models have enabled state-of-the-art results across a broad range of applications; however, training these models is extremely time- and resource-intensive, taking weeks on clusters with thousands of expensive accelerators in the extreme case. In this talk, I will describe two systems that improve the resource efficiency of model training. The first system, PipeDream, proposes the use of pipelining to accelerate distributed training. Pipeline parallelism facilitates model training with lower communication overhead than previous methods while still ensuring high compute resource utilization. Pipeline parallelism also enables the efficient training of large models that do not fit on a single worker. Pipeline parallelism is being used at Facebook, Microsoft, OpenAI, and Nvidia for efficient large-scale model training. The second system, Gavel, determines how resources in a shared cluster with heterogeneous compute resources (e.g., different types of hardware accelerators) should be partitioned among different users to optimize objectives specified over multiple training jobs. Gavel can improve various scheduling objectives, such as average completion time, makespan, or cloud computing resource cost, by up to 3.5x. I will conclude the talk with discussion on future directions for optimizing Machine Learning systems.

Bio: Deepak Narayanan is a final-year PhD student at Stanford University advised by Prof. Matei Zaharia. He is interested in designing and building software to improve the runtime performance and efficiency of emerging machine learning and data analytics workloads on modern hardware. His work is supported by a NSF graduate fellowship.


This talk will be recorded.  To request accommodations for a disability please contact Emily Lawrence, emilyl@cs.princeton.edu, at least one week prior to the event.

Integrating Machine Learning into Algorithm Design

Date and Time
Monday, February 22, 2021 - 12:30pm to 1:30pm
Location
Zoom Webinar (off campus)
Type
CS Department Colloquium Series
Host
Sanjeev Arora

Please register here.


Ellen Vitercik
An important property of those algorithms that are typically used in practice is broad applicability—the ability to solve problems across diverse domains. However, the default, out-of-the-box performance of these algorithms can be unsatisfactory, with slow runtime, poor solution quality, and even negative long-term social ramifications. In practice, there is often ample data available about the types of problems that an algorithm will be run on, data that can potentially be harnessed to fine-tune the algorithm’s performance. We therefore need principled approaches for using this data to obtain strong application-specific performance guarantees.

In this talk, I will give an overview of my research that provides practical methods built on firm theoretical foundations for incorporating machine learning and optimization into the process of algorithm design, selection, and configuration. I will describe my contributions across several diverse domains, including integer programming, clustering, mechanism design, and computational biology. As I will demonstrate, these seemingly disparate areas are connected by overarching structure which implies broadly-applicable guarantees.

Bio: Ellen Vitercik is a PhD student at Carnegie Mellon University where she is co-advised by Maria-Florina Balcan and Tuomas Sandholm. Her research revolves around artificial intelligence, algorithm design, and the interface between economics and computation, with a particular focus on machine learning theory. Among other honors, she is a recipient of the Exemplary Artificial Intelligence Track Paper Award at EC'19, the Best Presentation by a Student or Postdoctoral Researcher Award at EC'19, the NSF Graduate Research Fellowship, the IBM PhD Fellowship, the Fellowship in Digital Health from CMU's Center for Machine Learning and Health, and the Teaching Assistant of the Year Award from CMU's Machine Learning Department.


This talk will be recorded.  To request accommodations for a disability please contact Emily Lawrence, emilyl@cs.princeton.edu, at least one week prior to the event.

Exploiting Latent Structure and Bisimulation Metrics for Better Generalization in Reinforcement Learning

Date and Time
Monday, March 8, 2021 - 4:30pm to 5:30pm
Location
Zoom Webinar (off campus)
Type
CS Department Colloquium Series
Host
Tom Griffiths

Please register here.


Amy Zhang
The advent of deep learning has shepherded unprecedented progress in various fields of machine learning. Despite recent advances in deep reinforcement learning (RL) algorithms, however, there is no method today that exhibits anywhere near the generalization that we have seen in computer vision and NLP. Indeed, one might ask whether deep RL algorithms are even capable of the kind of generalization that is needed for open-world environments.  This challenge is fundamental and will not be solved with incremental algorithmic advances. 

In this talk, we propose to incorporate different assumptions that better reflect the real world and allow the design of novel algorithms with theoretical guarantees to address this fundamental problem. We first present how state abstractions can accelerate reinforcement learning from rich observations, such as images, without relying either on domain knowledge or pixel-reconstruction. Our goal is to learn state abstractions that both provide for effective downstream control and invariance to task-irrelevant details. We use bisimulation metrics to quantify behavioral similarity between states, and learn robust latent representations which encode only the task-relevant information from observations. We provide theoretical guarantees for the learned approximate abstraction and extend this notion to families of tasks with varying dynamics.

Bio: I am a final year PhD candidate at McGill University and the Mila Institute, co-supervised by Profs. Joelle Pineau and Doina Precup. I am also a researcher at Facebook AI Research. My work focuses on bridging theory and practice through learning approximate state abstractions and learning representations for generalization in reinforcement learning. I previously obtained an M.Eng. in EECS and dual B.Sci. degrees in Mathematics and EECS from MIT.


To request accommodations for a disability please contact Emily Lawrence, emilyl@cs.princeton.edu, at least one week prior to the event.

The Measurement and Mismeasurement of Trustworthy ML

Date and Time
Monday, March 1, 2021 - 12:30pm to 1:30pm
Location
Zoom Webinar (off campus)
Type
CS Department Colloquium Series
Host
Barbara Engelhardt

Please register here.


Sanmi Koyejo
Across healthcare, science, and engineering, we increasingly employ machine learning (ML) to automate decision-making that, in turn, affects our lives in profound ways. However, ML can fail, with significant and long-lasting consequences. Reliably measuring such failures is the first step towards building robust and trustworthy learning machines. Consider algorithmic fairness, where widely-deployed fairness metrics can exacerbate group disparities and result in discriminatory outcomes. Moreover, existing metrics are often incompatible. Hence, selecting fairness metrics is an open problem. Measurement is also crucial for robustness, particularly in federated learning with error-prone devices. Here, once again, models constructed using well-accepted robustness metrics can fail. Across ML applications, the dire consequences of mismeasurement are a recurring theme. This talk will outline emerging strategies for addressing the measurement gap in ML and how this impacts trustworthiness.

Bio: Sanmi (Oluwasanmi) Koyejo is an Assistant Professor in the Department of Computer Science at the University of Illinois at Urbana-Champaign. Koyejo's research interests are in developing the principles and practice of trustworthy machine learning. Additionally, Koyejo focuses on applications to neuroscience and healthcare. Koyejo has been the recipient of several awards, including a best paper award from the conference on uncertainty in artificial intelligence (UAI), a Kavli Fellowship, an IJCAI early career spotlight, and a trainee award from the Organization for Human Brain Mapping (OHBM). Koyejo serves on the board of the Black in AI organization.


To request accommodations for a disability please contact Emily Lawrence, emilyl@cs.princeton.edu, at least one week prior to the event.

Improving Stack-Wide Resource Utilization for a Faster Mobile Web

Date and Time
Monday, March 23, 2020 - 12:30pm to 1:30pm
Location
Zoom (off campus)
Type
CS Department Colloquium Series
Host
Karthik Narasimhan

***Due to the developing situation surrounding the COVID-19 virus, this talk will be available for remote viewing.  See below for details.***

Ravi Netravali
Abstract: Mobile web pages are integral to today's society, supporting critical services such as education, e-commerce, and social networking. Despite considerable academic and industrial research efforts, and major improvements over the past decade across the client-side web stack (i.e., networks, device CPUs, and browser engines), page load performance has plateaued and continues to fall short of user performance demands in practice. The consequences of this are far reaching: users abandon pages early, costing content providers billions of dollars in lost revenue; or pages are unusably slow, particularly in developing regions where web pages are often the sole gateway to the aforementioned services.

In this talk, I will describe the origin of this performance plateau in the context of serialized page load tasks that preclude effective utilization of the underlying network and CPU resources. Then, I will describe two complementary optimizations that my students and I have developed to eliminate these inefficiencies throughout the page load process and cut mobile load times in half. Key to these optimizations are judicious applications of programming languages (e.g., symbolic execution) and machine learning (e.g., reinforcement learning) techniques that enable us to 1) discover optimization knobs that preserve application correctness, and 2) tune those knobs according to stack-wide signals from the network, device, page, and browser, without developer intervention. I will conclude by describing how these underlying techniques can motivate and address a range of future challenges in networked applications and distributed systems. 

Bio: Ravi Netravali is an Assistant Professor of Computer Science at UCLA. His research interests are broadly in computer systems and networking, with a recent focus on building practical systems to improve the performance and debugging of large-scale, distributed applications for both end users and developers. His research has been recognized with an NSF CAREER Award, a Google Faculty Research Award, an ACM SoCC Best Paper Award, and an IRTF Applied Networking Research Prize. Prior to joining UCLA, Netravali received a PhD in Computer Science from MIT in 2018.


Zoom information:
Topic: Ravi Netravali CS Seminar
Time: Mar 23, 2020 12:00 PM Eastern Time (US and Canada)

Join Zoom Meeting
https://princeton.zoom.us/j/645162020

Meeting ID: 645 162 020

One tap mobile
+13126266799,,645162020# US (Chicago)
+16465588656,,645162020# US (New York)

Dial by your location
+1 312 626 6799 US (Chicago)
+1 646 558 8656 US (New York)
+1 253 215 8782 US
+1 301 715 8592 US
+1 346 248 7799 US (Houston)
+1 669 900 6833 US (San Jose)
Meeting ID: 645 162 020
Find your local number: https://princeton.zoom.us/u/avcvlf1F3

Join by SIP
645162020@zoomcrc.com

Join by H.323
162.255.37.11 (US West)
162.255.36.11 (US East)
221.122.88.195 (China)
115.114.131.7 (India Mumbai)
115.114.115.7 (India Hyderabad)
213.19.144.110 (EMEA)
103.122.166.55 (Australia)
209.9.211.110 (Hong Kong)
64.211.144.160 (Brazil)
69.174.57.160 (Canada)
207.226.132.110 (Japan)
Meeting ID: 645 162 020

Deep Probabilistic Graphical Modeling

Date and Time
Thursday, March 12, 2020 - 12:30pm to 1:30pm
Location
Zoom (off campus)
Type
CS Department Colloquium Series
Host
Ryan Adams

***Due to the developing coronavirus situation, this talk will now be available for remote viewing via Zoom.  See below for full details.***

Adji Bousso Dieng
Abstract: Deep learning (DL) is a powerful approach to modeling complex and large scale data. However, DL models lack interpretable quantities and calibrated uncertainty. In contrast, probabilistic graphical modeling (PGM) provides a framework for formulating an interpretable generative process of data and a way to express uncertainty about what we do not know. How can we develop machine learning methods that bring together the expressivity of DL with the interpretability and calibration of PGM to build flexible models endowed with an interpretable latent structure that can be fit efficiently? I call this line of research deep probabilistic graphical modeling (DPGM). In this talk, I will discuss my work on developing DPGM both on the modeling and algorithmic fronts. In the first part of the talk I will show how DPGM enables learning document representations that are highly predictive of sentiment without requiring supervision. In the second part of the talk I will describe entropy-regularized adversarial learning, a scalable and generic algorithm for fitting DPGMs. 

Bio: Adji Bousso Dieng is a PhD Candidate at Columbia University where she is jointly advised by David Blei and John Paisley. Her research is in Artificial Intelligence and Statistics, bridging probabilistic graphical models and deep learning. Dieng is supported by a Dean Fellowship from Columbia University. She won a Microsoft Azure Research Award and a Google PhD Fellowship in Machine Learning. She was recognized as a rising star in machine learning by the University of Maryland.  Prior to Columbia, Dieng worked as a Junior Professional Associate at the World Bank. She did her undergraduate studies in France where she attended Lycee Henri IV and Telecom ParisTech--France's Grandes Ecoles system. She spent the third year of Telecom ParisTech's curriculum at Cornell University where she earned a Master in Statistics.


Topic: Adji Bousso Dieng CS Seminar
Time: Mar 12, 2020 12:30 PM Eastern Time (US and Canada)

Join Zoom Meeting
https://princeton.zoom.us/j/384273957 

Meeting ID: 384 273 957

One tap mobile
+16465588656,,384273957# US (New York)
+16699006833,,384273957# US (San Jose)

Dial by your location
        +1 646 558 8656 US (New York)
        +1 669 900 6833 US (San Jose)
Meeting ID: 384 273 957
Find your local number: https://princeton.zoom.us/u/abUHt2KPwU

Join by SIP
384273957@zoomcrc.com

Join by H.323
162.255.37.11 (US West)
162.255.36.11 (US East)
221.122.88.195 (China)
115.114.131.7 (India Mumbai)
115.114.115.7 (India Hyderabad)
213.19.144.110 (EMEA)
103.122.166.55 (Australia)
209.9.211.110 (Hong Kong)
64.211.144.160 (Brazil)
69.174.57.160 (Canada)
207.226.132.110 (Japan)
Meeting ID: 384 273 957

Sharing without Showing: Building Secure Collaborative Systems

Date and Time
Tuesday, March 10, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Amit Levy

***Due to the developing coronavirus situation, we want to reduce the number of attendees at invited talks this week.  Attendance at this talk will now be limited to "Princeton University faculty only".  For other people who want to watch the talk, it will be available by livestream only, via the Princeton Broadcast Center for individual viewing.  We will not be hosting a separate room for the livestream.***

 

Wenting Zheng
In many domains such as finance and medicine, organizations have encountered obstacles in data acquisition because their target applications need sensitive data that reside across multiple parties. However, such data cannot be shared today due to data privacy concerns, policy regulation, and business competition.

My graduate research focused on solving this problem by enabling organizations to run complex computations on the joint dataset without revealing their sensitive input to the other parties. My overall approach is to co-design systems with cryptography to build practical and functional systems that provide strong and provable security. In this talk, I will focus on two systems — Opaque and Helen — which secure SQL analytics and machine learning, respectively. My open source has been used by organizations such as IBM Research, Ericsson, Alibaba, and Microsoft.

Bio: Wenting Zheng is a Ph.D. candidate at UC Berkeley, co-advised by Raluca Ada Popa and Ion Stoica. She completed her bachelor’s and master of engineering at MIT, where she was advised by Barbara Liskov. Her research interests are in computer systems, security, and applied cryptography. She is the recipient of a Berkeley Fellowship from 2014-2016, an IBM Research fellowship from 2017-2018, and was an invited participant at the 2019 EECS Rising Stars workshop. 

Designing Algorithms for Social Good

Date and Time
Monday, March 9, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series

Rediet Abebe
Algorithmic and artificial intelligence techniques show immense potential to deepen our understanding of socioeconomic inequality and inform interventions designed to improve access to opportunity. Interventions aimed at historically under-served communities are made particularly challenging by the fact that disadvantage and inequality are multifaceted, notoriously difficult to measure, and reinforced by feedback loops in underlying structures.

In this talk, we develop and analyze algorithmic and computational techniques to address these issues through two types of interventions: one in the form of allocating scarce societal resources and another in the form of improving access to information. We examine the ways in which techniques from algorithms, discrete optimization, and network and computational science can combat different forms of disadvantage, including susceptibility to income shocks, disparities in access to health information, and social segregation. We discuss current policy and practice informed by this work and close with a discussion of an emerging research area -- Mechanism Design for Social Good (MD4SG) -- around the use of algorithms, optimization, and mechanism design to address this category of problems.

Bio: Rediet Abebe is a Junior Fellow at the Harvard Society of Fellows. She holds a Ph.D. in computer science from Cornell University, where she was advised by Jon Kleinberg, as well as an M.S. in applied mathematics from Harvard University, an M.A. in mathematics from the University of Cambridge, and a B.A. in mathematics from Harvard College. Her research is in the fields of algorithms and AI, with a focus on discrete algorithms, optimization, network and computational science, and their applications to equity and social good concerns. As part of this research agenda, Abebe co-founded Mechanism Design for Social Good (MD4SG), a multi-institutional, interdisciplinary initiative working to improve access to opportunity. This initiative has participants from over 100 institutions in 20 countries and has been supported by Schmidt Futures, the MacArthur Foundation, and the Institute for New Economic Thinking.

Abebe's work has informed policy and practice at various organizations, including the Ethiopian Ministry of Education and the National Institutes of Health. In 2019, she served on the NIH Advisory Committee to the Director Working Group on AI, whose recommendations were unanimously approved by the General Director's advisory committee. Abebe was recently recognized by the 2019 MIT Technology Review’s 35 Innovators Under 35 award and honored as a one to watch by the 2018 Bloomberg 50 list. She has presented her research in venues such as the National Academy of Sciences, the United Nations, and the Museum of Modern Art. Her work has been covered by outlets including Forbes, the Boston Globe, and the Washington Post. In 2017 Abebe co-founded Black in AI, a non-profit organization tackling diversity and inclusion issues in the field. Her research is deeply influenced by her upbringing in her hometown of Addis Ababa, Ethiopia.

This talk is being co-sponsored by CITP and the Department of Computer Science.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Software-Hardware systems for the Internet of Things

Date and Time
Thursday, March 5, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Kyle Jamieson

Omid Abari
Recently, there has been a huge interest in Internet of Things (IoT) systems, which bring the digital world into the physical world around us. However, barriers still remain to realizing the dream applications of IoT. One of the biggest challenges in building IoT systems is the huge diversity of their demands and constraints (size, energy, latency, throughput, etc.). For example, virtual reality and gaming applications require multiple gigabits-per-second throughput and millisecond latency. Tiny sensors spread around a greenhouse or smart home must be low-cost and batteryless to be sustainable in the long run. Today's networking technologies fall short in supporting these IoT applications with a hugely diverse set of constraints and demands. As such, they require distinct innovative solutions.

In this talk, I will describe how we can design a new class of networking technologies for IoT by designing software and hardware jointly, with an understanding of the intended application. In particular, I will present two examples of our solutions. The first solution tackles the throughput limitations of existing IoT networks by developing new millimeter wave devices and protocols, enabling many new IoT applications, such as untethered high-quality virtual reality. The second solution tackles the energy imitations of IoT networks by introducing new wireless devices that can sense and communicate without requiring any batteries. I demonstrate how our solution is applicable in multiple, diverse domains such as HCI, medical, and agriculture. I will conclude the talk with future directions in IoT research, both in terms of technologies and applications.

Bio: Omid Abari is an Assistant Professor at the University of Waterloo, School of Computer Science. He received his Ph.D. in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology (MIT) in 2018. His research interests are in the area of computer networks and mobile systems, with applications to the Internet of Things (IoT). He is currently leading the Intelligent Connectivity (ICON) Lab, where his team focuses on the design and implementation of novel softwarehardware systems that deliver ubiquitous sensing, communication and computing at scale. His work has been selected for GetMobile research highlights (2018, 2019), and been featured by several media outlets, including Wired, TechCrunch, Engadget, IEEE Spectrum, and ACM Tech News.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

The Value Alignment Problem in Artificial Intelligence

Date and Time
Monday, February 24, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Tom Griffiths

Dylan Hadfield-Menell
Much of our success in artificial intelligence stems from the adoption of a simple paradigm: specify an objective or goal, and then use optimization algorithms to identify a behavior (or predictor) that optimally achieves this goal. This has been true since the early days of AI (e.g., search algorithms such as A* that aim to find the optimal path to a goal state), and this paradigm is common to AI, statistics, control theory, operations research, and economics. Loosely speaking, the field has evaluated the intelligence of an AI system by how efficiently and effectively it optimizes for its objective. This talk will provide an overview of my thesis work, which proposes and explores the consequences of a simple, but consequential, shift in perspective: we should measure the intelligence of an AI system by its ability to optimize for our objectives.

In an ideal world, these measurements would be the same -- all we have to do is write down the correct objective! This is easier said than done: misalignment between the behavior a system designer actually wants and the behavior incentivized by the reward or loss functions they specify is routine, it is commonly observed in a wide variety of practical applications, and fundamental, as a consequence of limited human cognitive capacity. This talk will build up a formal model of this value alignment problem as a cooperative human-robot interaction: an assistance game of partial information between a human principal and an autonomous agent. It will begin with a discussion of a simple instantiation of this game where the human designer takes one action, write down a proxy objective, and the robot attempts to optimize for the true objective by treating the observed proxy as evidence about the intended goal. Next, I will generalize this model to introduce Cooperative Inverse Reinforcement Learning, a general and formal model of this assistance game, and discuss the design of efficient algorithms to solve it. The talk will conclude with a discussion of directions for further research including applications to content recommendation and home robotics, the development of reliable and robust design environments for AI objectives, and the theoretical study of AI regulation by society as a value alignment problem with multiple human principals.

Bio: Dylan is a final year Ph.D. student at UC Berkeley, advised by Anca Dragan, Pieter Abbeel, and Stuart Russell. His research focuses on the value alignment problem in artificial intelligence. His goal is to design algorithms that learn about and pursue the intended goal of their users, designers, and society in general. His recent work has focused on algorithms for human-robot interaction with unknown preferences and reliability engineering for learning systems.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Follow us: Facebook Twitter Linkedin