Speaker: Steven Xia, Ph.D. Candidate, University of Illinois Urbana-Champaign

Abstract:

AI and large language models (LLMs) are ushering in a new era of software development, fundamentally reshaping how systems are built, tested, and maintained. In this talk, I will present my research on integrating modern LLM capabilities with core software engineering principles to automate complex development tasks and strengthen the coding and reasoning abilities of AI systems. I will first introduce TitanFuzz, which leverages LLMs to implicitly solve complex and heterogeneous constraints for large-scale system testing, uncovering critical bugs in real-world software, and inspiring extensive subsequent work in LLM-driven testing. Next, I will discuss Agentless, which demonstrates that a simple, lightweight framework can effectively solve challenging repository-level software engineering tasks without heavy agentic overhead. I will showcase how Agentless has been adopted by leading AI companies to evaluate their latest models, inspire new agentic frameworks, and perform post- or mid-training for agentic coding and reasoning. In addition, I will highlight my other work on automating complex software tasks and my recent work to enable runtime self-improvement in AI coding systems. Together, my work establishes a scalable and practical foundation for building more capable and reliable AI-driven software systems.

Bio:

Chunqiu Steven Xia is a PhD candidate in Computer Science at UIUC, advised by Professor Lingming Zhang. His research lies at the intersection of Software Engineering and Artificial Intelligence, with a focus on automating critical and cognitively demanding software development tasks using large language models (LLMs) and software agents. His work has significantly advanced how fundamental software problems are addressed in both academia and industry, uncovering 200+ critical bugs and vulnerabilities in widely used systems, and being adopted by leading AI companies, including OpenAI, Meta, DeepSeek, and MiniMax. His research has been recognized with an ACM SIGSOFT Distinguished Paper Award (FSE 2025), an Oral Spotlight Paper Award (COLM 2024), a CACM Research Highlight (FSE 2023), and the prestigious Amazon AICE PhD Fellowship.

Speaker:  Shanto Rahman, Ph.D. Candidate, University of Texas at Austin 

Abstract: 

Software testing is one of the most scalable ways to check program correctness. However, nondeterminism in test execution and rapid code evolution can cause tests to break and produce false failures, where the same test may fail without any underlying bug. As these false failures accumulate, developers may start ignoring test results, allowing real bugs to reach production. My research aims to make software testing reliable by detecting, understanding, reproducing, and repairing such broken tests. In this talk, I first present my work on predicting and diagnosing the root causes of flaky tests using context-aware attribution, helping developers understand why nondeterministic failures occur while reducing runtime and memory cost. Next, I introduce automated repair techniques for both nondeterminism- and evolution-induced test breakage using intent-preserving dynamic instrumentation and generative AI. These techniques are evaluated on large-scale, real-world datasets, and achieve high repair success. I conclude by outlining a path from reliable test diagnosis and repair, extending these reliability foundations to modern computing systems for trustworthy, safe operation in practice. 

Bio: 

Shanto Rahman is a final-year Ph.D. candidate at The University of Texas at Austin, advised by August Shi. Her research sits at the intersection of software engineering, program analysis, and AI, with a focus on making software testing reliable under nondeterminism and continuous code evolution. Her work has been published in top venues including ICSE, OOPSLA, ASE, and ICST. Shanto has gained industry experience through research internships at Google and Amazon Web Services (AWS). She has been recognized as an MIT EECS Rising Star and has received multiple UT Austin fellowships and awards. 

Speaker:  Sourav Chakraborty, Professor, Indian Statistical Institute

Abstract: 

The problem of testing whether an input satisfies a given property can be significantly easier than estimating the input’s distance from having that property. However, for certain classes of properties, these tasks may be quite similar.

One of the newer models for testing distribution properties is the huge object model, introduced by Goldreich and Ron in 2023. In this talk, we investigate the query complexity of testing index-invariant properties within this model. We adapt Szemerédi’s regularity method to this setting and prove that, for index-invariant properties, constant query testability implies constant-query estimability.

This talk is based on the following works:

1. “Testing vs Estimation for Index-Invariant Properties in the Huge Object

Model,” joint work with Eldar Fischer, Arijit Ghosh, Amit Levi, Gopinath Mishra,

and Sayantan Sen, STOC 2025.

2. “Testing of Index-Invariant Properties in the Huge Object Model,” joint work

with Eldar Fischer, Arijit Ghosh, Gopinath Mishra, and Sayantan Sen, COLT 2023

Bio:

Sourav Chakraborty is a Professor in the Advanced Computing and Microelectronics Unit (ACMU) of the Computer and Communication Sciences Division (CCSD) at the Indian Statistical Institute (ISI), Kolkata, India. Before joining ISI in July 2018, he was a faculty member at the Chennai Mathematical Institute, India, from September 2010. He completed his PhD in Computer Science in June 2008 at the University of Chicago under the supervision of Prof. László Babai and did postdoctoral stints at Technion, Israel, and CWI, Amsterdam.

Speaker:  Courtney Miller, Ph.D. Candidate, Carnegie Mellon University 

Abstract: 

Software development is shaped by constant disruption—from open-source dependency abandonment to generative AI (GenAI) tools fundamentally reshaping how code is written. 
 
In this talk, I'll present two lines of work, each examining a different disruption. First, I'll share my work on the disruption of open source dependency abandonment, revealing the unsupported challenges developers face when identifying and responding to abandonment, identifying factors that support successful downstream response, and demonstrating how theory-driven LLM-based systems can make those risks more visible. Second, I’ll present findings from my recent collaboration with Microsoft Research about the disruption of GenAI adoption on development teams, demonstrating why adoption succeeds for some developers but stalls for others in seemingly similar contexts, showing the amplification effect organizational support mechanisms can have, and introducing the Productivity Pressure Paradox: a dynamic where increased productivity expectations without corresponding support can actually prevent developers from building the skills that would deliver those gains. 
 
Finally, I’ll outline my research vision for helping developers anticipate and navigate the disruptions of tomorrow—from building AI-powered tools for software supply chain resilience to studying how GenAI is reshaping development team practices and processes. 

Bio:

Courtney Miller is a Software Engineering Ph.D. Candidate in the School of Computer Science at Carnegie Mellon University, co-advised by Bogdan Vasilescu and Christian Kästner. Her research examines how developers navigate technological disruptions—from dependency abandonment threatening open source software supply chain sustainability and security to generative AI adoption reshaping development workflows. 
 
Her empirical approach combines human-centered qualitative techniques with large-scale data-driven analysis, modeling, and visualization to inform the development of theoretically-grounded insights and practice-driven tooling solutions. As part of the Secure Software Supply Chain Center (S3C2), her work has informed security practices through contributions to federal agency and industry supply chain security summits. 
 
Courtney has earned three ACM SIGSOFT Distinguished Paper Awards at the International Conference on Software Engineering (ICSE) and an NSF Graduate Research Fellowship. She has conducted two research internships at Microsoft Research. She holds a B.A. with honors in Computer Science and Statistics from New College of Florida. In her free time, she enjoys indoor cycling and leisurely walks with her 14-year-old toy poodle Chanel. 

 1 p.m. "Modernizing C4G BLIS: The Road to BLIS 4.0"

Mitchell Rysavy (Senior Software Engineer, Microsoft)

 1:30 p.m."Computing for Good in an Increasingly Bad World"

Michael Best (Professor, School of International Affairs and School of Interactive Computing, Georgia Tech)

 2 p.m. "Disaster Response in Rural Communities"

Justin McLellan (Software Engineering Manager, Crexi)

 2:30 p.m. "HandS together: Innovation and Collaboration for Greater Impact"

Protip Biswas (Vice President, United Way Regional Commission on Homelessness)

Date and Time: Nov. 13 at 11 a.m.

Location: KACB 2443

Title: Distributed Optimization and Consensus in the Presence of Adversarial Agents 

Abstract: Due to their many potential applications, there is a rich history of research on distributed consensus and distributed optimization. Distributed consensus requires the agents in the system to reach agreement on an output as a function of the inputs of the various agents. Distributed optimization requires the agents to collaboratively optimize a global cost function that consists of the summation of the local cost functions of the individual agents. Loss minimization in machine learning is a special case of this problem, and it has many other applications as well. The problems of consensus and optimization are inter-related, with some commonalities in the solutions used for these problems.

In this talk, we will consider the above two problems in the presence of adversarial agents. The adversarial agents may misbehave and attempt to tamper with the outcome of the distributed computation. Such adversarial behavior may occur due to security compromise of the agents or due to software/hardware failure. We will take an overview of some recent results on the above two problems, with an intuitive presentation of the proposed solutions.

The talk will NOT assume any distributed computing background from the audience.

Bio: Nitin Vaidya is the Robert L. McDevitt, K.S.G., K.C.H.S. and Catherine H. McDevitt L.C.H.S. Chair Professor Computer Science at Georgetown University, where he served as the Department Chair during 2018-24. His current research interests are in the area of distributed algorithms, and previously he has worked on wireless networks. He received a Ph.D. from the University of Massachusetts at Amherst. He previously served as a Professor and Associate Head in Electrical and Computer Engineering at the University of Illinois at Urbana-Champaign. He has co-authored papers that received awards at several conferences, including SSS, ACM MobiHoc and ACM MobiCom. He is a fellow of the IEEE. He has served as the Chair of the Steering Committee for the ACM PODC conference, as the Editor-in-Chief for the IEEE Transactions on Mobile Computing, and as the Editor-in-Chief for ACM SIGMOBILE publication MC2R.

Title: Neural-symbolic Software Auditing

Abstract: Software engineering is undergoing a major shift: while code generation has become increasingly automated, code review and audit remain stubbornly human-intensive. Code quality continues to be a persistent challenge, and developers often face the task of debugging or auditing code they did not write. LLMs hold promise for automating aspects of code analysis, yet they consistently fall short in auditing real-world repositories due to context limitations, hallucinations, and difficulty with repository-scale reasoning. In this talk, I will present our recent research that addresses these challenges. Specifically, RepoAudi is an autonomous, LLM-driven auditing agent designed for repository-level code analysis with high precision and efficiency. By mimicking expert auditors, it performs demand-driven, path-sensitive reasoning over control- and data-flow graphs, powered by abstraction, pointer tracking, and validation mechanisms. This approach has enabled RepoAudit to uncover hundreds of previously unknown bugs in mature software ecosystems, including the Linux kernel and OpenSSL. In order to extend the tool to audit binary executables that do not have any source code information, we developed a novel code-model training method and award-winning probabilistic program analysis approaches. These techniques enable RepoAudit to find dozens of vulnerabilities in real-world firmware by solely performing auditing. At the end, I will briefly discuss how to audit future software, namely, AI agents.

Bio: Xiangyu Zhang is a Samuel Conte professor at Purdue specializing in Software Engineering, AI red-teaming, and Cyber Forensics. His work involves developing techniques to detect bugs, including security vulnerabilities, in traditional software systems as well as AI models and systems, and to leverage AI techniques to perform software engineering and cybersecurity tasks. He has served as the Principal Investigator (PI) for numerous projects funded by organizations such as DARPA, IARPA, ONR, NSF, AirForce, and industry.

• Register for the 5K: https://gatech.co1.qualtrics.com/jfe/form/SV_3EMUEFBWUSdMLS6?Q_CHL=qr
• Volunteer to support the 5K: https://gatech.co1.qualtrics.com/jfe/form/SV_eVsLxvaYtIl3TaC

Join us as we learn strategies for acknowledging, valuing, and respecting differences in the workplace and leave with action steps for building communication around diversity, equity, and inclusion.  Lunch will be served. 

Event:  Diversity Activity and Q&A with Cedric Stallworth
Date:  Tuesday, October 18
Time:  11 a.m. - 12:30 p.m.
Place:  Klaus Bldg. Room 1116W

To send your RSVP and submit your question for the Q&A session with Cedric Stallworth, visit https://gatech.co1.qualtrics.com/jfe/form/SV_cDaZCUdpLuJ6DPw.

The success of deep neural networks (DNNs) from Machine Learning (ML) has inspired domain specific architectures (DSAs) for them. Google’s first generation DSA offered 50x improvement over conventional architectures for ML inference in 2015. Google next built the first production DSA supercomputer for the much harder problem of training. Subsequent generations greatly improved performance of both phases. We start with ten lessons learned from such efforts. 

The rapid growth of DNNs rightfully raised concerns about their carbon footprint. The second part of the talk identifies the “4Ms” (Model, Machine, Mechanization, Map) that, if optimized, can reduce ML training energy by up to 100x and carbon emissions up to 1000x. By improving the 4Ms, ML held steady at <15% of Google’s total energy use despite it consuming ~75% of its floating point operations.  With continuing focus on the 4Ms, we can realize the amazing potential of ML to positively impact many fields in a sustainable way.

David Patterson is a UC Berkeley Pardee professor emeritus, a Google distinguished engineer, and the RISC-V International Vice-Chair. His most influential Berkeley projects likely were RISC and RAID. His best-known book is Computer Architecture: A Quantitative Approach. He and his co-author John Hennessy shared the 2017 ACM A.M Turing Award and the 2022 NAE Charles Stark Draper Prize for Engineering. The Turing Award is often referred to as the “Nobel Prize of Computing” and the Draper Prize is considered a “Nobel Prize of Engineering.”

Where: GT Hotel Club Room
800 Spring St NW, Atlanta, GA 30308
the Club Room is located on the first floor of the Hotel
When: Friday, October 21, 5 - 8 p.m.

There will be pool tables, a raffle, and a photo booth at this event!

Refreshments and drink tickets provided. So take a break from your studies, and come relax and have fun!

Please RSVP by October 18

https://art.c21u.gatech.edu/

Where: GT Hotel Club Room
800 Spring St NW, Atlanta, GA 30308
the Club Room is located on the first floor of the Hotel
When: Thursday, March 31, 5 - 8 p.m.

There are two pool tables in the club room and we have reserved a photo booth so you can have a souvenir of the evening. Refreshments and drink tickets provided. So take a break from your studies, and come relax and have fun!

March 11 
11 a.m. - 2 p.m. 
In-person and virtual participation welcome. Register here to save your spot! https://tinyurl.com/WomenInDataScience22 

The College of Computing and the Machine Learning Center at Georgia Tech are hosting Dean, who is speaking on March 8, in the Clough Auditorium (144). His presentation, Five Exciting Trends in Machine Learning, begins at 11 a.m.

Power of Two Schedule

9 -10:30 a.m.| Breakfast for everyone in Computing!... bacon - Klaus Atrium 

11 a.m. -12 p.m. | Computing Student Conversations w/ the Dean - Klaus 1443 

1200pm-130pm | GT Computing Faculty and Staff Lunch - Klaus Atrium

2 -4 p.m. | Afternoon snacks and games for everyone in Computing - Klaus 1116 

2:22 p.m. | Giveaways - Klaus 1116  

5:00 p.m. | In-person Fireside Chat with Dean Isbell - Clough 144
(Join virtually https://primetime.bluejeans.com/a2m/live-event/fcqeauxz)

Live stream for Computing, Alumni and OMSCS  

6:30 – 9 p.m. | Reception and Monte Carlo Night - Alumni, Students, Faculty, and Staff - Klaus Atrium and Room 1116

The faculty recruiting seminar is scheduled for Feb. 8 at 12 p.m. in the 9th floor atrium of the Coda Building.

Abstract: Blockchains are an exciting area of research that touches on many areas of Computer Science and beyond. This technology has the potential to enable a fast, cheap, and private financial system based on distributed consensus and cryptography, instead of trusted parties. Despite this potential, the reality still shows severe limitations of blockchains: (i) transactions can cost hundreds of dollars and take minutes to confirm, (ii) some blockchains offer little privacy, and (iii) proof-of-work consensus consumes too much energy. In this talk, I will discuss powerful techniques that follow a prover paradigm and can mitigate these limitations. The first technique, called Bulletproofs, is a general-purpose zero-knowledge proof system that is specifically designed to enable confidential blockchain transactions. Bulletproofs requires minimal trust assumptions and gives the shortest zero-knowledge proofs without a trusted setup. The system is widely deployed and powers tens of thousands of private blockchain transactions per day. The second technique, called inner pairing products, is a way to aggregate many zero-knowledge proofs into a single short proof. This can significantly reduce on-chain data, leading to a significant increase in transactions per second that the chain can process. The third technique is a new concept called a verifiable delay function (VDF) that is vital for permission-less and eco-friendly consensus. VDFs are already deployed in Filecoin and Chia, and are planned for Ethereum 2.0, the upcoming upgrade to Ethereum.

TITLE:  Old and New Challenges in Networking

BIO: 

Dr. Ellen Zegura is Regents’ and Fleming Endowed Professor of Computer Science at Georgia Tech. She works in two primary areas, computer networking and computing for social good. In computer networking, she is known for her work on Internet topology tools and for inventing the message ferries for communicating in sparse networks. Her work in computing and social good includes work in Liberia with the Carter Center, with Native Americans in the Southwest, and with residents of the Westside of Atlanta. In education she is co-creator of the GT Serve Learn Sustain initiative, which prepares students to use their disciplinary skills to create sustainable communities. She co-led the Civic Data Science summer internship program, an NSF-funded Research Experience for Undergraduates site that has trained more than 70 students. She is a Fellow of the IEEE and ACM, and recently stepped down as Chair of the Computing Research Association Board.

ABSTRACT:

Data science has provided unprecedented opportunities to learn new insights and to predict, recommend, cluster, classify, transform, and optimize. Catalyzed by large-scale, networked computer systems, vast availability of data, and machine learning algorithms, data science has been extraordinarily impactful to-date, and it holds great promise in all disciplines. However, no new technology arrives without complications, and we have recently seen both the press and various political circles illustrating real, potential, and fictional implications of the field.

This presentation aims to balance the opportunities provided by data science against the many challenges that have ensued.  At its core, the talk proposes a rubric that practitioners can apply to tease out data science’s complexities and also maps out seven categories of data sciences challenges, ranging from engineering to ethics. The talk is illustrated with examples from many applications, and it concludes with some suggested ways to address the downsides of the field.

BIO:

Dr. Alfred Spector was most recently CTO and Head of Engineering at Two Sigma, a firm dedicated to algorithmic approaches to a wide collection of financial optimization problems. His career has led him from innovation in large scale, networked computing systems to broad engineering and research leadership.  Most recently, he has been writing a book on Data Science, which will be published in early 2022.  
 

Prior to Two Sigma, Spector spent nearly eight years as VP of Research and Special Initiatives at Google. Before Google, Spector held various senior-level positions at IBM, including as global VP of Services and Software Research and global CTO of IBM’s Software Business. He previously founded Transarc Corporation, a pioneer in distributed transaction processing and wide-area file systems (which was acquired by IBM), and he was a tenured professor at Carnegie Mellon University.  Dr. Spector obtained a Ph.D. in computer science from Stanford and a B.A. in applied math from Harvard.

Spector was a Hertz Fellow at Stanford and is also a Fellow of both the ACM and the IEEE. He is an active member of the National Academy of Engineering and the American Academy of Arts and Sciences, where he serves on the Council. Dr. Spector won the 2001 IEEE Kanai Award for Distributed Computingand the 2016 ACM Software Systems Award. In 2018-19, Dr. Spector lectured widely as a Phi Beta Kappa Scholar and has been a member of the ACM Turing Award Committee.  As to government service,  Spector was a member of the Army Science Board, and he chaired the NSF’s CISE Advisory Board.  He has had extensive international experience due to broad responsibilities at IBM, Google, and Two Sigma.

Recently, Spector has lectured widely on the growing importance of computer science across all disciplines based on the evocative phrase, CS+X.  More recently, he has written and lectured on the societal implications of data science -- both the great benefits and the unintended consequences. He has participated in many committees on education and technology-related policy.

[Based on joint work with Jingling Chen, Nan Jiang, Sham Kakade, Akshay Krishnamurthy, Aditya Modi and Wen Sun]

Bio: Alekh Agarwal is a researcher who works on theoretical foundations of machine learning, spanning many areas including large-scale and distributed optimization, high-dimensional statistics, online learning, and most recently reinforcement learning. He focuses on designing theoretically sound methods which lend themselves to practice, and his work at Microsoft has resulted in the creation of a new Azure service (https://aka.ms/personalizer) that operationalizes some of his reinforcement learning research. His work has been recognized with a NeurIPS best paper award and an ACM SIGAI Industry Impact award for his work on the Azure Personalization Service.

Register and Attend: https://primetime.bluejeans.com/a2m/register/qxugrckz

Bio: Francis Bach is a researcher at INRIA in the Computer Science department of Ecole Normale Supérieure, in Paris, France. He has been working on machine learning since 2000, with a focus on algorithmic and theoretical contributions, in particular in optimization. Past papers can be downloaded from his web page or Google Scholar page.

Register and Attend: https://primetime.bluejeans.com/a2m/register/eqttretz

Twenty years later, those same assistant professors ended up in the College of Computing at Georgia Tech to continue their lifelong collaboration. One of them, W. Storey Professor of Computing Richard “Dick” Lipton was the first Georgia Tech computer scientist elected to the National Academy of Engineering. The other was Richard “Rich” DeMillo, who left his position as Chief Technology Officer at Hewlett-Packard to become the John P. Imlay Jr. Dean of Computing.

Their seminal 1979 work, “Social Processes and Proofs of Theorems and Programs,” is now part of a new collection of 46 classic papers in computer science published as a book this year from the MIT Press. Ideas That Created the Future by Harvard Professor Harry Lewis spans the intellectual birth and growth of the field – from Leibnitz and Boole to Knuth and RSA – and covers the sweeping discoveries and advancements that have come to define computer science. As Lewis points out in his introductory essay, “The fact that some computer scientists still bristle when this paper is mentioned is testament to its dialectic force.”

The College of Computing at Georgia Tech invites you to take part in a historical conversation between Rich DeMillo and Dick Lipton, two pioneers in computer science who helped shape a field that has come to reshape how we live every day.

Join the virtual fireside chat, Thursday, June 17, 2021, at 7 PM ET, for this living history – as told by the men who lived it – and hear about the experiences and their roles in this turning point in computer science.

Students, families, faculty and staff, please join us for a virtual Dean's Welcome New Graduate Alumni Celebration, to be held, May 15 at 12 - 2 p.m. EST on Gatherly (links and additional information still being determined)

The program is as follows;

• Welcome: Jennifer Whitlow, Director of Computing Enrollment and Alumni Engagement
• Charles Isbell, Dean of the College of Computing, John P. Imlay Jr. Chair
• David Joyner, Executive Director, Online Education & OMSCS
• Closing: Jennifer Whitlow

Total run time is ~2 hours

Be sure to tag your Dean’s New Alumni Celebration photos and posts on Instagram and Twitter with #GTComputing and #GT21

We wish all of our graduates a safe, healthy, and enjoyable summer!

Title: Developing Robust Models, Algorithms, Databases, and Tools with Applications to Cybersecurity and Healthcare 

Date: Wednesday, May 12, 2021

Time: 12pm-2pm EST 

Location (virtual): https://bluejeans.com/8164507038/

Scott Freitas

Machine Learning Ph.D. Student

School of Computational Science and Engineering
Georgia Institute of Technology

https://www.scottfreitas.com

Committee

• Duen Horng (Polo) Chau [Advisor, Associate Professor, CSE, Georgia Institute of Technology]
• Srijan Kumar [Assistant Professor, CSE, Georgia Institute of Technology]
• Diyi Yang [Assistant Professor, CSE, Georgia Institute of Technology]

Abstract

As society and technology becomes increasingly interconnected, so does the threat landscape. Once isolated threats now pose serious concerns to highly interdependent systems, highlighting the fundamental need for robust machine learning. This dissertation contributes novel tools, algorithms, databases and models—through the lens of robust machine learning—in a research effort to solve large-scale societal problems affecting millions of people in the areas of cybersecurity and healthcare.

1. Tools: We develop TIGER, the first comprehensive graph robustness toolbox; and our Robustness Survey identifies critical yet missing areas of graph robustness research.

2. Algorithms: Our survey and toolbox reveal existing work has overlooked lateral attacks on computer authentication networks. We develop D2M, the first algorithmic framework to quantify and mitigate network vulnerability to lateral attacks by modeling lateral attack movement from a graph theoretic perspective.

3. Databases: To prevent lateral attacks altogether, we develop MalNet-Graph, the world’s largest cybersecurity graph database—containing over 1.2M graphs across 696 classes—and show the first large-scale results demonstrating the effectiveness of malware detection through a graph medium. We plan to extend MalNet-Graph by constructing the largest binary-image cybersecurity database—containing 1.2M images, 133x more images than the only other public database—enabling new discoveries in malware detection and classification research restricted to a few industry labs (MalNet-Image).

4. Models: To protect systems from adversarial attacks, we develop UnMask, the first model that flag semantic incoherence in computer vision systems, which detects up to 96.75% of attacks, and defends the model by correctly classifying up to 93% of attacks. Inspired by UnMask's ability to protect computer visions systems from adversarial attack, we develop REST, which creates noise robust models through a novel combination of adversarial training, spectral regularization and sparsity regularization. In the presence of noise, our method improves state-of-the-art sleep stage scoring by 71%--allowing us to diagnose sleep disorders earlier on and in the home environment—while using 19x less parameters and 15x less MFLOPS. 

ABSTRACT:

Deep learning has fundamentally revolutionized our society, including recommendation systems that suggest movies and items to purchase, virtual assistants, and  future autonomous vehicles. From a theoretical point of view, there are two main questions in deep learning: (1) understanding of why deep learning works, (2) how to speed the running time of deep learning algorithms. In this talk, I will provide progress for these two questions. First, I will show that stochastic gradient descent is able to memorize all the input data points under some mild assumptions. Second, I will present a new optimization algorithm for fast deep learning. The key observation is that stochastic gradient descent is an iterative algorithm, and gradients change slowly in each iteration. This observation allows us to update gradients in a lazy fashion, significantly improving per-iteration running time. Finally, I will briefly describe how to apply similar ideas to speed up convex optimization and quantum machine learning.

BIO:

 Zhao Song is a postdoctoral researcher at Princeton University and Institute for Advanced Study. He obtained his Ph.D. from the Computer Science Department at University of Texas at Austin under the supervision of Professor Eric Price in 2019. His Ph.D. thesis won Bert Kay Dissertation Award (the best dissertation award) at the Computer Science Department in the University of Texas at Austin. His recent work, InstaHide, won second place for the Bell Lab Prize, which is awarded to the top-3 most innovative technology breakthroughs annually.

JOIN TALK HERE: https://bluejeans.com/946032411  

Registration is required

Multiclass and Multi-Label Learning with General Losses: What is the Right Output Coding and Decoding?

Abstract

Many practical applications of machine learning involve multiclass learning problems with a large number of classes -- indeed, multi-label learning problems can be viewed as a special case. Multiclass learning with the standard 0-1 loss is fairly well understood; however, in practice, applications with large numbers of classes often require performance to be measured via a different, problem-specific loss. What is the right way to design principled and efficient learning algorithms for multiclass (and multi-label) problems with general losses? 

From a theoretical standpoint, an elegant approach for designing statistically consistent learning algorithms is via the design of convex calibrated surrogate losses. From a practical standpoint, an approach that is often favored is that of output coding, which reduces multiclass learning to a set of simpler binary classification problems. In this talk, I will discuss recent progress in bringing together these seemingly disparate approaches under a unifying lens to develop statistically consistent and computationally efficient learning algorithms for a wide range of problems, in some cases recovering existing state-of-the-art algorithms, and in other cases providing new ones. Our algorithms require learning at most r real-valued scoring functions, where r is the rank of the target loss matrix, and come with corresponding principled decoding schemes. I will also discuss connections with the field of property elicitation, and new tools for deriving quantitative regret transfer bounds via strongly proper losses.

About Shivani

Shivani Agarwal is Rachleff Family Associate Professor of Computer and Information Science at the University of Pennsylvania, where she also directs the NSF-sponsored Penn Institute for Foundations of Data Science (PIFODS) and co-directs the Penn Research in Machine Learning (PRiML) center. She is currently an Action Editor for the Journal of Machine Learning Research and an Associate Editor for the Harvard Data Science Review. She has previously been a Radcliffe Fellow at Harvard University, an Assistant Professor and Ramanujan Fellow at the Indian Institute of Science, and a postdoctoral lecturer at MIT. She received her PhD in computer science at the University of Illinois, Urbana-Champaign, and a bachelors degree in computer science as a Nehru Scholar at Trinity College, University of Cambridge. Her research interests include foundational questions in machine learning, applications of machine learning in the life sciences, and connections between machine learning and other disciplines such as economics, operations research, and psychology. More broadly, she is excited by research at the intersection of computer science, mathematics, and statistics, and its ability to turn data into actionable insights in both the natural and social sciences.

Registration is required. Register here.

Generative models based on point processes for financial time series simulation

Abstract: 

In this seminar, I will talk about generative models based on point processes for financial time series simulation. Specifically, we focus on a recently developed state-dependent Hawkes (sdHawkes) process to model the limit order book dynamics [Morariu-Patrichi, 2018]. The sdHawkes model consists of an oracle Hawkes process and a state process following Markov transition. The Hawkes and state processes are fully coupled, which enables the point process captures the self-and cross-excitation as well as the interaction between events and states. We will go through the model formulation in sdHawkes, the simulation of sdHawkes, its maximum likelihood estimation, and more importantly, its application to high-frequency data modeling that captures the interactions between the order flow and the state of the current market.

Morariu-Patrichi, Maxime, and Mikko S. Pakkanen. "State-dependent Hawkes processes and their application to limit order book modelling." arXiv preprint arXiv:1809.08060 (2018).

About Qi:

Qi Wei received his Ph.D. degree in machine learning and image processing from the National Polytechnic Institute of Toulouse (INPENSEEIHT), University of Toulouse, France in September 2015, and Bachelor degree in Electrical Engineering from Beihang University (BUAA), Beijing, China in July 2010. Wei's doctoral thesis Bayesian Fusion of Multi-band Images: A Powerful Tool for Super-resolution was rated as one of the best theses (awarded Prix Leopold Escande) at the University of Toulouse, 2015.

Wei has worked on multiband image processing as a Research Associate with Signal Processing Laboratory, University of Cambridge, UK, and as a Research Associate at Duke University, US. He has also worked at Siemens Corporate Technology as a Research Scientist. Since 2018, Wei served as a vice president and machine learning scientist at JPMorgan. His research has been focused on machine/deep learning, time series analysis, computer vision/image processing, Bayesian statistical inference, etc.

ABSTRACT:

BIO:

Marios Kogias is a researcher at MSR Cambridge. He graduated from EPFL in August 2020. His main research focus is at the intersection of operating systems and networking in the datacenter. He’s working on building and understanding systems with strict tail-latency SLOs leveraging new emerging hardware. He was an IBM PhD Fellow and won the best student paper award at Eurosys2020. Before joining EPFL he got his undergrad degree from the National Technical University of Athens. He has interned at Cern, Google, and Microsoft Research

JOIN THE TALK: https://bluejeans.com/946032411

ABSTRACT:

Machine learning is becoming widely used in decision making, in domains ranging from personalized medicine and mobile health to online education and recommendation systems. While (supervised) machine learning traditionally excels at prediction problems, decision making requires answering questions that are counterfactual in nature, and ignoring this mismatch leads to unreliable decisions. As a consequence, our understanding of the algorithmic foundations for data-driven decision making is limited, and efficient algorithms are typically developed on an ad hoc basis. Can we bridge this gap and make decision making as easy as machine learning? Focusing on the contextual bandit, a core problem in data-driven decision making, we bridge the gap by providing the first optimal and efficient reduction to supervised machine learning. The algorithm allows users to seamlessly apply off-the-shelf supervised learning models and methods to make decisions on the fly, and has been implemented in widely-used, industry-standard tools for decision making. Our results advance a broader program to develop a universal algorithm design paradigm for data-driven decision making. I will close the talk by discussing challenges and opportunities in building such a framework, including efforts to extend our developments to difficult reinforcement learning problems in large state spaces.
 

BIO:

Dylan Foster is a postdoctoral fellow at the MIT Institute for Foundations of Data Science. He holds a Ph.D. in computer science from Cornell University, where he was advised by Karthik Sridharan. He has received several awards, including the best paper award at COLT (2019), best student paper award at COLT (2018, 2019), Facebook Ph.D. fellowship, and NDSEG PhD fellowship. His research focuses on problems at the intersection of learning and decision making.

JOIN TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Current hardware and software systems were conceived at a time when we had scarce compute and memory resources, limited quantity of data and users, and easy hardware performance scaling due to Moore’s Law. These assumptions are not true today. Today, emerging web services require data centers that scale to hundreds of thousands of servers, i.e., hyperscale, to efficiently process requests from billions of users. In this new era of hyperscale computing, we can no longer afford to build each layer of the systems stack separately. Instead, we must rethink the synergy between the software and hardware worlds from the ground up.

 In this talk, I will focus on re-thinking (1) software threading and concurrency paradigms and (2) data center hardware architectures. First, I will describe μTune, my software threading framework that is aware of the overheads induced by the underlying hardware’s constraints. Then, I will discuss SoftSKU and Accelerometer—my proposals to answer the question of: How should we build data center hardware for emerging software paradigms in the post-Moore era? Finally, I will conclude by describing my ongoing and future research towards re-designing the systems stack to enable the hyperscale web services of tomorrow.

BIO:

Akshitha Sriraman is a Ph.D. candidate in Computer Science and Engineering at the University of Michigan. Her research bridges computer architecture and software systems, demonstrating the importance of that bridge in realizing efficient hyperscale web services via solutions that span the systems stack. Her systems solutions to improve hardware efficiency have been deployed in real hyperscale data centers and currently serve billions of users, saving millions of dollars and meaningfully reducing the global carbon footprint. Additionally, her hardware design proposals have influenced Intel’s Alder Lake+ CPU architectures.

Sriraman has been recognized with a Facebook Fellowship, a Rackham Merit Ph.D. Fellowship, and was selected for the Rising Stars in EECS Workshop. Her work has been recognized with an IEEE Micro Top Picks distinction and has appeared in top architecture and systems venues like OSDI, ISCA, ASPLOS, MICRO, and HPCA.

JOIN THIS TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Despite notable successes on several carefully controlled benchmarks, current machine learning (ML) systems are remarkably brittle, raising serious concerns about their deployment in safety-critical applications like self-driving cars and predictive healthcare. In this talk, I discuss fundamental obstacles to building robust ML systems and develop principled approaches that form the foundations of robust ML. In particular, I will focus on the role of data and demonstrate the need to question common assumptions when improving robustness to (i) adversarial examples and (ii) spurious correlations. On the one hand, I will describe how and why naively using more data can surprisingly hurt performance in these settings. On the other hand, I will show that unlabeled data, when harnessed in the right fashion, is extremely beneficial and enables state-of-the-art robustness. In closing, I will discuss how to build on the foundations of robust ML and achieve wide-ranging robustness in various domains including natural language processing and vision.

BIO:

Aditi Raghunathan is a fifth-year Ph.D. student at Stanford University advised by Percy Liang. She is interested in building robust machine learning systems with guarantees for trustworthy real-world deployment. Her research in robustness has been recognized by a Google Ph.D. Fellowship in Machine Learning and the Open Philanthropy AI Fellowship. Among other honors, she is also the recipient of the Anita Borg Memorial Scholarship and the Stanford School of Engineering Fellowship.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

With the proliferation of distributed systems, the design of safe, secure, and efficient software has become an ever more complex task. The heterogeneous nature of these distributed systems have further introduced domain-specific programming requirements such as programming in an adversarial setting and inferring execution cost. Recognizing these requirements, my research has focused on advancing programming languages to support developers with program verification techniques, complexity analysis tools, and domain-specific languages.

In this talk, I focus on the domain of smart contracts, i.e., programs that enforce digital transactions, often legal or financial, between multiple parties. Programming smart contracts comes with its unique challenges, which include enforcing protocols of interaction, analyzing execution cost, and tracking linear assets. To address these challenges, the talk introduces Nomos, a programming language for smart contracts. To express and enforce protocols, Nomos is based on session types rooted in linear logic. To predict execution cost, Nomos uses resource-aware session types and automatic amortized resource analysis, a type-based technique for inferring cost bounds. To track assets, Nomos employs a linear type system that prevents assets from being duplicated or discarded.

The talk concludes with my future plans on exploring how programming languages can aid in the synthesis of smart contracts, verification of cryptographic protocols, and analysis of probabilistic systems with applications in artificial intelligence.

BIO:

Ankush Das is a final-year Ph.D. student at Carnegie Mellon University. He is advised by Professor Jan Hoffmann and closely works with Professor Frank Pfenning. He is broadly interested in programming languages with a specific focus on program verification, complexity analysis, and domain-specific languages. He is the lead designer and developer of Nomos, a programming language for smart contracts rooted in resource-aware session types. He has also designed the Rast language based on refinement session types that won the best system description paper award at FSCD 2020. Before joining CMU, he worked as a Research Fellow at Microsoft Research, India where he designed static analysis and verification tools for Windows driver modules. He completed his undergraduate at IIT Bombay, India in 2014 where he worked on deciding termination of linear loop programs.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Software bugs are pervasive in modern software. As software is integrated into increasingly many aspects of our lives, these bugs have increasingly severe consequences, both from a security (e.g. Cloudbleed, Heartbleed, Shellshock) and cost standpoint. Fuzzing refers to a set of techniques that automatically find bug-triggering inputs by sending many random-looking inputs to the program under test. In this talk, I will discuss how, by identifying core under-generalized components of modern fuzzing algorithms, and building algorithms that generalize or tune these components, I have expanded the application domains of fuzzing. First, by building a general feedback-directed fuzzing algorithm, I enabled fuzzing to consistently find performance and resource consumption errors. Second, by developing techniques to maintain structure during mutation, I brought fuzzing exploration to “deeper” program states. Third, by decoupling the user-facing abstraction of random input generators from their sampling distributions, I built faster validity fuzzing and even tackled program synthesis. Finally, I will discuss the key research problems that must be tackled to make fuzzing readily-available and useful to all developers.

BIO:

Caroline Lemieux is a final-year Ph.D. candidate at UC Berkeley, advised by Koushik Sen. Her research aims to help developers improve the correctness, security, and performance of large, existing software systems, ranging from complex open-source projects to industrial-scale software. Her current projects tackle these goals with a focus on fuzz testing and program synthesis. Her work on fuzz testing has been awarded an ACM SIGSOFT Distinguished Paper Award, ACM SIGSOFT Distinguished Artifact Award, ACM SIGSOFT Tool Demonstration Award, and Best Paper Award (Industry Track). Before Berkeley, she received her B.Sc. in combined honours computer science and mathematics at the University of British Columbia, where she won the Governor General’s Silver Medal in Science (highest standing in the Faculty of Science). She is the recipient of a Berkeley Fellowship for Graduate Study and a Google Ph.D. Fellowship in Programming Technologies and Software Engineering.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Data is eating the world. Systems for storing and processing data, such as database
management systems (DBMSs), are thus pivotal for our computing infrastructure. It is critical
that they function correctly — incorrectly computed results (e.g., by omitting a row) can cause
serious loss or damage. Despite their importance, finding such logic bugs in production DBMSs is a longstanding challenge. This talk presents novel, general approaches to effectively detecting logic bugs in DBMSs by tackling the test oracle problem, i.e., deciding whether the returned result for a query is correct. These approaches were realized as the SQLancer open-source tool, which was evaluated on widely-used, production-quality DBMSs, such as SQLite, MySQL, and PostgreSQL. To date, SQLancer has found over 450 unique previously unknown bugs in these systems, most of which have been fixed by the developers. This body of work has provided solid methodological and technical bases for effectively testing DBMSs and has already been widely adopted by industry. Many reliability challenges remain, not only for DBMSs, but also for other data-centric systems. The heterogeneous landscape in this space provides an exciting, fertile ground for new, practically relevant research for improving the reliability and performance of our society’s data processing infrastructure

BIO:

Manuel Rigger is a postdoctoral researcher in the Advanced Software Technologies (AST)
Lab at ETH Zurich, mentored by Zhendong Su. His research goal is to make software more
reliable by developing practical and principled approaches to tackle important challenges related to correctness, performance, and security. He draws on and contributes to the software engineering, systems, programming languages, and compilers fields. His recent work has focused on automatically testing database management systems, and led to over 450 unique bugs in widely-used systems such as SQLite, MySQL, PostgreSQL, TiDB, and CockroachDB being discovered, most of which have been fixed. He completed his Ph.D. at Johannes Kepler University Linz, mentored by Hanspeter Mössenböck, and worked on the safe and efficient execution of unsafe languages (project Sulong). Part of this work was integrated into Oracle’s GraalVM to support its execution of LLVM IR.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Fueled by evolving applications, requirements from systems and networks are changing in fundamental ways: networks today must provide increasingly strict performance guarantees and not just high performance. However, the design principles of the early Internet continue to be the basis of modern networks to provide best-effort performance. This leads to a mismatch between what applications need and what networks provide. To bridge this gap, we need predictability to be built into the design of networks.

In this talk, I will focus on the unique challenges and unprecedented opportunities for building predictable networks. I will first discuss this in the context of shared public clouds by presenting a new abstraction and system, PicNIC, for predictable network performance while navigating fundamental trade-offs. Then, I will expand the discussion with wide-area networks with a system to achieve efficiency and predictability. Finally, I will highlight research directions towards leveraging hardware support to build predictable networks in a scalable, robust, and efficient manner.

BIO:

Praveen Kumar is a Ph.D. candidate in computer science at Cornell University. His research focuses on building predictable networks and takes a multi-disciplinary approach spanning architecture, PL, systems, and theory to build holistic networking solutions — from low-level hardware to high-level abstractions and algorithms. Kumar’s systems have been deployed at scale and also influenced the design of several other systems in production. His work has been recognized with a SIGCOMM best student paper award.

JOIN THE TALK HERE: https://bluejeans.com/946032411

Registration is required.

You can lead a horse to water...: Representing vs. Using Features in Neural NLP
 

Abstract

A wave of recent work has sought to understand how pretrained language models work. Such analyses have resulted in two seemingly contradictory sets of results. On one hand, work based on "probing classifiers" generally suggests that SOTA language models contain rich information about linguistic structure (e.g., parts of speech, syntax, semantic roles). On the other hand, work which measures performance on linguistic "challenge sets" shows that models consistently fail to use this information when making predictions. In this talk, I will present a series of results that attempt to bridge this gap. Our recent experiments suggest that the disconnect is not due to catastrophic forgetting nor is it (entirely) explained by insufficient training data. Rather, it is best explained in terms of how "accessible" features are to the model following pretraining, where "accessibility" can be quantified using an information-theoretic interpretation of probing classifiers.
 

About Ellie

Ellie Pavlick is an Assistant Professor of Computer Science at Brown University where she leads the Language Understanding and Representation (LUNAR) Lab. She received her PhD from the one-and-only University of Pennsylvania. Her current work focuses on building more cognitively-plausible models of natural language semantics, focusing on grounded language learning and on sample efficiency and generalization of neural language models.

ABSTRACT:

Distributed systems are increasingly important for our everyday life, allowing for high performance, fault tolerance, and flexibility. Many of these systems nowadays rely on the inherently insecure internet infrastructure. Surely though, they should have been designed to take this into account?

In this talk, I will answer negatively to this question using a concrete example: public blockchain systems such as Bitcoin. These are novel distributed systems that are designed according to stringent failure models. In this context, I will explain how an adversary controlling pieces of Internet infrastructure can practically compromise: (i) Bitcoin’s consensus protocol (by partitioning the network); (ii) Bitcoin’s anonymity guarantees (by mapping pseudonyms to real-world identities); and (iii) Bitcoin’s availability (by eclipsing clients).

While these attacks are worrying, I will also introduce practical and effective defenses to counter them both at the network and the application layer. Beyond Bitcoin, this work teaches essential lessons for distributed-system design.

BIO:

Maria Apostolaki is a Ph.D. student at ETH Zurich advised by Laurent Vanbever.
During her studies, she has been a visiting student at MIT (2019) and a research intern at Microsoft Research (2018) and Google (2017). Before joining ETH, she earned her diploma in electrical and computer engineering at the National Technical University of Athens, Greece.

Her research focuses on building secure, performance, and deployable networked systems using both hardware and software. Her research led to discovering significant vulnerabilities in the Bitcoin system and spearheaded changes to the Bitcoin codebase. Apostolaki’s work received widespread media coverage and was awarded an Applied Networking Research Prize by IETF.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Modern algorithms have to regularly deal with uncertain inputs. This uncertainty can take many forms, e.g., in online advertisement future users are unknown (the input arrives online), in spectrum-auctions bidder valuations are unknown (the users are strategic), and in oil-drilling the amount of oil is unknown (the input is stochastic). Traditionally, there has not been significant overlap in the study of these different forms of uncertainty. I believe that studying these uncertainties together gives us a lot more power. In this talk, I will give an overview of my research in “Algorithms and Uncertainty” where I am able to successfully use these relationships.

Studying these different forms of uncertainty together allows us to find connections and build unifying techniques. As an example, I will talk about my progress on long-standing combinatorial auctions problems that deal with strategic inputs, by using techniques which were originally developed for online inputs. Moreover, a combined study of uncertainty helps us find richer cross-cutting models. For example, several important online problems do not admit good algorithms in the classical worst-case models. I will talk about how to give a “beyond the worst-case” analysis for such problems and obtain more nuanced performance guarantees, by using models/techniques arising in other forms of uncertainty.

BIO:

Sahil Singla is a research instructor (postdoc) jointly between Princeton University and Institute for Advanced Study. He received his Ph.D. in computer science from Carnegie Mellon University, where he was advised by Anupam Gupta and Manuel Blum. His research is in algorithms and uncertainty where the goal is to design optimal algorithms for uncertain inputs by studying different forms of uncertainty together. Singla has served on the program committees of the conferences SODA, ICALP, EC, and ESA. His research has been invited for talks at Highlights Beyond EC, at China Theory Week, at TCS+, and at Highlights of Algorithms. He recently contributed a chapter to the book Beyond the Worst-Case Analysis of Algorithms.

JOIN THE TALK HERE: https://bluejeans.com/946032411

ABSTRACT:

Decision making is central to a broad range of domains, including finance, transportation, healthcare, the travel industry, robotics, and engineering. It is often found at the very final step of business analytics--prescriptive analytics--to allow businesses to transform a rich understanding of data, typically provided by advanced predictive models, into actionable decisions. Modeling and solving these problems have relied on application-specific solutions, which are often complex, error-prone, and not generalizable. My goal is to create a domain-independent, declarative approach, supported and powered by the system where the data relevant to these problems typically resides: the database. Despite the widespread importance of prescriptive analytics, unified solutions close to the data did not exist.In my talk, I will present a prototype system that supports package queries, a new query modelthat extends traditional database queries to handle complex constraints and preferences over answer sets. Package queries allow the declarative specification and efficient evaluation of a significant class of constrained optimization problems--integer programs--within a database. These queries pose unique challenges to a database system, ranging from their richer expressive power, more complex semantics, and harder computational complexity than their SQL counterpart, to scalability issues that arise from large amounts of data and uncertainty in the data. I will illustrate how our unified system addresses all these challenges achieving high performance and quality in many real-world problems from finance, healthcare, and science. I will also present my vision for data-centric systems for decision making, and their connections with robotics, machine learning, natural language processing, visualization, operations management, and simulation.

BIO:

Matteo Brucato is a Ph.D. candidate in computer science at the University of Massachusetts Amherst. Matteo's research aims at augmenting data management systems to better support data science and all stages of analytics, with a focus on prescriptive analytics and data-driven decision making. He is the co-inventor of package queries and his work has been recognized by multiple awards, including CACM and SIGMOD Record research highlights, best paper at VLDB 2016, and both the best demonstration and the runner-up award in VLDB 2020. Matteo's work is largely interdisciplinary, spanning multiple research areas beyond data management, such as natural language processing, information retrieval, AI, machine learning, robotics, and operations research. He received his Bachelor's and Master's degrees in computer science from the University of Bologna (Italy), he visited Aarhus University (Denmark), UC Riverside, UC Berkeley, and NYU, and he interned at MSR and IBM.

JOIN THE TALK HERE: https://bluejeans.com/946032411

Registration is required

Hardness of MDP planning with linear function approximation

Markov decision processes (MDPs) is a minimalist framework to capture that many tasks require long-term plans and feedback due to noisy dynamics. Yet, as a result MDPs lack structure and as such planning and learning in MDPs with the typically enormous state and action spaces is strongly intractable; no algorithm can avoid Bellman's curse of dimensionality in the worst case. However, as recognized already by Bellman and his co-workers at the advent of our field, for many problem of practical interest, the optimal value function of an MDP is well approximated by just using a few basis functions, such as those that are standardly used in numerical calculations. As knowing the optimal value function is essentially equivalent to knowing how to act optimally, one hopes that this observation can be turned into efficient algorithms as there are only a few coefficients to compute. If this is possible, we can think of the resulting algorithms as performing computations with a compressed form of the value functions. While many algorithms have been proposed as early as in the 1960s, until recently not much has been known about whether these compressed computations are possible and when. In this talk, I will discuss a few recent results (some positive, some negative) that are concerned with these compressed computations and conclude with some open problems. As we shall see, still today, there are more open questions than questions that have been satisfactorily answered.

About Csaba

Csaba Szepesvari is a Canada CIFAR AI Chair, the team-lead for the “Foundations” team at DeepMind and a Professor of Computing Science at the University of Alberta. He earned his PhD in 1999 from Jozsef Attila University, in Szeged, Hungary. He has authored three books and over 200 peer-reviewed journal and conference papers. He serves as the action editor of the Journal of Machine Learning Research and Machine Learning, as well as on various program committees. Dr. Szepesvari's interest is artificial intelligence (AI) and, in particular, principled approaches to AI that use machine learning. He is the co-inventor of UCT, a widely successful Monte-Carlo tree search algorithm. UCT ignited much work in AI, such as DeepMind's AlphaGo which defeated the top Go professional Lee Sedol in a landmark game. This work on UCT won the 2016 test-of-time award at ECML/PKDD.

ABSTRACT:

BIO:

Aayush Jain is a final year Ph.D. student, advised by Amit Sahai at UCLA. He is interested in research questions in cryptography and their wonderful connections with related areas in computer science.

His most notable contribution is the first construction of Program Obfuscation whose security rests upon four well-studied hard problems.  This problem was widely considered to be one of the most consequential open problems in cryptography. His work on this problem has been the subject of an invited tutorial at FOCS 2020, an article in Quanta magazine, and a Simons Institute Workshop. His research is recognized by a Google PhD Fellowship (2018-present), a Dean’s Fellowship (2016), and the 2020 Symantec Outstanding Graduate Research Award at UCLA.

JOIN THE TALK: https://bluejeans.com/821585375

ABSTRACT:

Data volumes are growing exponentially, fueled by an increased number of automated processes such as sensors and devices. Meanwhile, the computational power available for processing this data – as well as analysts’ ability to interpret it – remain limited. As a result, database systems must evolve to address these new bottlenecks in analytics. In my work, I ask: how can we adapt classic ideas from database query processing to modern compute- and analyst-limited data analytics?In this talk, I will discuss the potential for this kind of systems development through the lens of several practical systems I have developed. By drawing insights from database query optimization, such as pushing workload-  and domain-s pecific filtering, aggregation, and sampling into core analytics workflows, we can dramatically improve the efficiency of analytics at scale. I will illustrate these ideas by focusing on two systems — one designed for high-volume seismic waveform analysis and one designed to optimize visualizations for streaming infrastructure and application telemetry — both of which have been field-tested at scale. I will also discuss lessons from production deployments at companies including Datadog, Microsoft, Google and Facebook

BIO:

Kexin Rong is a Ph.D. student in computer science at Stanford University, co-advised by Professors Peter Bailis and Philip Levis. She designs and builds systems to enable data analytics at scale, supporting applications including scientific analysis, infrastructure monitoring, and analytical queries on big data clusters. Prior to Stanford, she received her bachelor’s degree in computer science from California Institute of Technology.

JOIN THE TALK HERE:  https://bluejeans.com/538930562

Time: 02/12/2021, 12.30pm - 1.30pm

Location: https://primetime.bluejeans.com/a2m/live-event/jzpdzxqz

Title: Addressing Biases for Robust, Generalizable AI

Abstract:

Artificial Intelligence has made unprecedented progress in the past decade. However, there still remains a large gap between the decision-making capabilities of humans and machines. In this talk, I will investigate two factors to explain why. First, I will discuss the presence of undesirable biases in datasets, which ultimately hurt generalization. I will then present bias mitigation algorithms that boost the ability of AI models to generalize to unseen data. Second, I will explore task-specific prior knowledge which aids robust generalization, but is often ignored when training modern AI architectures. Throughout this discussion, I will focus my attention on language applications, and will show how certain underlying structures can provide useful biases for inferring meaning in natural language. I will conclude with a discussion of how the broader framework of dataset and model biases will play a critical role in the societal impact of AI, going forward.

Bio:

Swabha Swayamdipta is a postdoctoral investigator at the Allen Institute for AI, working with Yejin Choi. Her research focuses on natural language processing, where she explores dataset and linguistic structural biases, and model interpretability. Swabha received her Ph.D. from Carnegie Mellon University, under the supervision of Noah A. Smith and Chris Dyer. During most of her Ph.D. she was a visiting student at the University of Washington. She holds a Masters degree from Columbia University, where she was advised by Owen Rambow. Her research has been published at leading NLP and machine learning conferences, and has received an honorable mention for the best paper at ACL 2020.

ABSTRACT:

One of the most successful applications of modern cryptography has been its use in electronic and digital payment systems. In traditional systems, a trusted authority handles all payments (e.g., a bank or a credit card company). More recently, blockchain systems have emerged as a trust-free and increasingly popular alternative.

In a blockchain system, users jointly emulate the trusted authority by running a distributed protocol to agree on the transaction history of users (i.e., the blockchain). Making blockchain systems a secure and scalable environment poses many new and fascinating challenges that require solutions from both cryptography and distributed computing. In my talk, I will explain the different areas of my research and their importance as components that make up a blockchain system. For each of these areas, I will also list some of the open questions that I plan to work on in the near future.

BIO:

Julian Loss obtained his MSc in computer science from ETH Zurich in 2016 and his Ph.D. from the Ruhr University of Bochum in 2019. He is currently a postdoc at the University of Maryland in the group of Jonathan Katz. His interests include classic cryptographic primitive,s such as digital signatures and multi-party computation, as well as blockchain/consensus protocols.

JOIN THE TALK HERE: https://bluejeans.com/821585375

ABSTRACT:

ABSTRACT:

Our society's dependence on the internet is unquestioned. Yet our ability to manage the risks associated with relying on the proper functioning of the internet's core infrastructure is hindered by a poor understanding of how it can fail. For example, we lack a thorough understanding of when/where/how/why large connectivity failures happen; or we rely on a global routing system that is extremely vulnerable and we do not know to what extent it is exploited, by whom, and for what purposes. My research tries to address these questions by acquiring empirical evidence (e.g., on network status and misuse) on a global scale and longitudinally. 

 In this talk I will first describe an approach I developed to gather unprecedented insight into episodes of connectivity disruption, which is based on two intuitions: the opportunity to extract (connectivity liveness) "signals" from so-called "Internet background noise" traffic and the benefit of combining different types of control-plane and data-plane measurements.  I then discuss how this inspired me to build a system (IODA,ioda.caida.org) to monitor the Internet for large-scale connectivity outages. IODA can detect Internet outages worldwide within minutes, continuously updating and monitoring hundreds of thousands of time series. IODA serves both as a research testbed and as a public service that provides automatically generated alerts and visualizations. This effort has enabled further research in multiple directions, facilitated interdisciplinary collaborations, and resulted in the engagement of industry, government, and civil society. 

 I will conclude my talk by briefly discussing my other ongoing and future research efforts to understand and remediate Internet infrastructural issues and weaknesses.

BIO:

Alberto Dainotti is an associate research scientist at CAIDA, the Center for Applied Internet Data Analysis at University of California, San Diego. His research is at the intersection of internet measurement, data science, and internet security with a focus on understanding and improving internet infrastructure security and reliability. He has received the ACM SIGCOMM Internet Measurement Conference Distinguished Paper Award, two IRTF Applied Networking Research Prizes, and the Best of ACM SIGCOMM Computer Communication Review award. Dainotti received his Ph.D. in computer engineering and systems at University of Napoli “Federico II,” Italy.  

Join the talk here: https://bluejeans.com/640828498

Registration is required

Title: Building the Next-Generation AI: Small and Efficient Neural Computing

Abstract:

Deep learning has changed the computing paradigm. Today, AI researchers & practitioners increasingly use deep neural networks for many applications across different modalities and areas such as NLP, Vision, Speech, Conversational and Multimodal AI. However, much of the Deep Learning revolution has been limited to the Cloud and highly specialized hardware. Recently the AI community has witnessed an increasing trend for training larger and larger neural models (e.g., GPT-3, T5, BERT) that achieve state-of-the-art results but require enormous computation, memory and energy resources on the Cloud. In order to enable AI experiences in real-time across all users and devices, ML models have to run efficiently on the Cloud and personal devices on the Edge (e.g., mobile phones, wearables, IoT) which have limited computing capabilities.

In this talk, I will introduce our work on Neural Projection computing, an efficient AI paradigm, and a family of efficient Projection Neural Network architectures that yield fast (e.g., quadratic speedup for transformer networks) and tiny models that shrink memory requirements by upto 10000x while achieving near state-of-the-art performance powering vision and NLP applications on billions of mobile devices. Widespread increase in availability of connected “smart” appliances (e.g., conversational assistants) means that there is an ever-expanding surface area for mobile intelligence and ambient devices in homes. Our approach enables efficient ML to solve complex prediction tasks for such applications both on-device and on Cloud, keeping model size, compute and power usage low while simultaneously optimizing for accuracy.

Bio:

Dr. Sujith Ravi is a Director at Amazon Alexa AI where he is leading efforts to build the future of multimodal conversational AI experiences at scale. Prior to that, he was leading and managing multiple ML and NLP teams and efforts in Google AI. He founded and headed Google’s large-scale graph-based semi-supervised learning platform, deep learning platform for structured and unstructured data as well as on-device machine learning efforts for products used by billions of people in Search, Ads, Assistant, Gmail, Photos, Android, Cloud and YouTube. These technologies power conversational AI (e.g., Smart Reply), Web and Image Search; On-Device predictions in Android and Assistant; and ML platforms like Neural Structured Learning in TensorFlow, Learn2Compress as Google Cloud service, TensorFlow Lite for edge devices.

Dr. Ravi has authored over 100 scientific publications and patents in top-tier machine learning and natural language processing conferences. His work has been featured in press: Wired, Forbes, Forrester, New York Times, TechCrunch, VentureBeat, Engadget, New Scientist, among others, and also won the SIGDIAL Best Paper Award in 2019 and ACM SIGKDD Best Research Paper Award in 2014. For multiple years, he was a mentor for Google Launchpad startups. Dr. Ravi was the Co-Chair (AI and deep learning) for the 2019 National Academy of Engineering (NAE) Frontiers of Engineering symposium. He was also the Co-Chair for EMNLP 2020, ICML 2019, NAACL 2019, and NeurIPS 2018 ML workshops and regularly serves as Senior/Area Chair and PC of top-tier machine learning and natural language processing conferences like NeurIPS, ICML, ACL, NAACL, AAAI, EMNLP, COLING, KDD, and WSDM.

Website: www.sravi.org 

Twitter: @ravisujith

LinkedIn: https://www.linkedin.com/in/sujithravi

Registration is required

TALK TITLE
Learning Tree Models in Noise: Exact Asymptotics and Robust Algorithms

ABSTRACT 

We consider the classical problem of learning tree-structured graphical models but with the twist that the observations are corrupted in independent noise. For the case in which the noise is identically distributed, we derive the exact asymptotics via the use of probabilistic tools from the theory of strong large deviations. Our results strictly improve those of Bresler and Karzand (2020) and Nikolakakis et al. (2019) and demonstrate keen agreement with experimental results for sample sizes as small as that in the hundreds. When the noise is non-identically distributed, Katiyar et al. (2020) showed that although the exact tree structure cannot be recovered, one can recover a "partial" tree structure; that is, one that belongs to the equivalence class containing the true tree. We propose Symmetrized Geometric Averaging (SGA), a statistically robust algorithm for partial tree recovery. We provide error exponent analyses and extensive numerical results on a variety of trees to show that the sample complexity of SGA is significantly better than the algorithm of Katiyar et al. (2020). SGA can be readily extended to Gaussian models and is shown via numerical experiments to be similarly superior.

https://arxiv.org/abs/2101.08917

https://arxiv.org/abs/2005.04354


This is joint work with Anshoo Tandon, Aldric J. Y. Han and Shiyao Zhu.

ABOUT VINCENT

Vincent Y. F. Tan received the B.A. and M.Eng. degrees in electrical and information sciences from Cambridge University and the Ph.D. degree in electrical engineering and computer science (EECS) from the Massachusetts Institute of Technology (MIT).

He is currently a Dean’s Chair Associate Professor with the Department  of Electrical and Computer Engineering and the Department of Mathematics, National University of Singapore (NUS). His research interests include information theory, machine learning, and statistical signal processing. 

He was also an IEEE Information Theory Society Distinguished Lecturer in 2018/9. He is currently serving as an Associate Editor for the IEEE Transactions on Signal Processing and an Associate Editor for Machine Learning for the IEEE Transactions on Information Theory. He is a member of the IEEE Information Theory Society Board of Governors.

ABSTRACT:

Matrix multiplication is one of the most basic algebraic operations. Since Strassen’s surprising breakthrough algorithm from 1969, which showed that matrices can be multiplied faster than the most straightforward algorithm, algorithmic problems from nearly every area of computer science have been sped up by clever reductions to matrix multiplication. It is popularly conjectured that two n × n matrices can be multiplied in roughly O(n^2) time, but we don’t yet have the algorithmic techniques needed to achieve this.

In this talk, I’ll give an overview of my work on this important problem in the broad context of my research on algorithm design and complexity theory for algebraic problems. First, I’ll give an overview of known applications of matrix multiplication, including some new applications to nearest neighbor search and processing data streams. Second, I’ll talk about the fastest known algorithm for matrix multiplication, which runs in time O(n^2.37286), which I recently designed with Virginia Vassilevska Williams. Third, I’ll describe new barrier results which help to explain why we’ve been stuck for so long on this problem, and describe what kinds of insights are needed for further improvements.

BIO:

Josh Alman is a Michael O. Rabin postdoctoral fellow in theoretical computer science at Harvard University. He earned his Ph.D. in computer science from MIT in 2019, where he was advised by Virginia Vassilevska Williams and Ryan Williams. He works on both algorithm design and complexity theory, and much of his work combines tools and insights from both areas. He’s particularly interested in developing algebraic tools like algorithms for matrix multiplication and Fourier transforms, and applying them to problems throughout computer science. His awards include the European Association of TCS Distinguished Dissertation Award, the George M. Sprowls Award for outstanding Ph.D. theses in computer science at MIT, and best student paper awards at CCC 2019 and FOCS 2019.

WATCH HERE: https://bluejeans.com/493104294

ABSTRACT:

Searching and summarization are two of the most fundamental tasks in massive data analysis. In this talk, I will focus on these two tasks from the perspective of diversity and fairness. Search is often formalized as the (approximate) nearest neighbor problem. Despite an extensive research on this topic, its basic formulation is insufficient for many applications. In this talk, I will describe such applications and our approaches to address them. For example, we show how to incorporate diversity or fairness in the results of a search query.

A prominent approach to summarize the data is to compute a small “core-set”: a subset of the data that is sufficient for approximating the solution of a given task. We introduce the notion of “composable core-sets” as core-sets with the composability property: the union of multiple core-sets should form a good summary for the union of the original data sets. This composability property enables efficient solutions to a wide variety of massive data processing applications, including distributed computation (e.g. Map-Reduce model), streaming algorithms, and similarity search. We show how to produce such efficient summaries of the data while preserving the diversity in the data set. I will describe several metrics for capturing the notion of diversity, and present efficient algorithms for construction of composable core-sets with respect to those metrics.

BIO:

Sepideh Mahabadi is a research assistant professor at the Toyota Technological Institute at Chicago (TTIC).  She received her Ph.D. from MIT, where she was advised by Piotr Indyk. For a year, she was a postdoctoral research scientist at Simons Collaboration on Algorithms and Geometry based at Columbia University. Her research focuses on theoretical foundations of massive data including high dimensional computational geometry, streaming algorithms, and data summarization; as well as social aspects of algorithms for massive data including diversity maximization and algorithmic fairness.

WATCH HERE:  https://bluejeans.com/457461626

Time: 01/29/2021, 12.30pm - 1.30pm

Location: https://primetime.bluejeans.com/a2m/live-event/eatygusw

Title: NLP in the Wild: From Ancient Akkadian to Biochemistry Protocols

Abstract:

I’ll present two recent projects showing the range of domains I’d like to tackle in my work to help experts with diverse real-world research questions. First, I’ll present a model capable of filling in missing parts in ancient cuneiform tablets written thousands of years ago in now-extinct languages (Akkadian and Sumerian). Due to deterioration over time, these excavated tablets are often broken, faded, or cracked, making it hard for historians and archaeologists to read and interpret them. We show that by leveraging large-scale language models pretrained on modern texts we achieve good results in restoring missing parts in various domains and time periods, in the automatic evaluation as well as human analysis. Second, I will discuss a novel document-level representation of wet lab biochemistry protocols geared towards experiment automation and reproducibility, addressing challenges such as cross-sentence relations, long-range coreference, grounding, and implicit arguments. I’ll show examples from a manually-annotated corpus of complex lab protocols, and present graph-prediction models that form the first step towards fully executable lab protocols.

Bio:

Dr. Gabriel Stanovsky is a senior lecturer (≈ assistant professor) at the Hebrew University of Jerusalem. He did his postdoctoral research at the University of Washington and the Allen Institute for AI in Seattle, working with Prof. Luke Zettlemoyer and Prof. Noah Smith, and his PhD with Prof. Ido Dagan at Bar-Ilan University. He is interested in developing NLP models that with benefits for users in real-world applications. His work has received awards at top-tier conferences, including ACL and CoNLL, and recognition in popular journals such as Science and The New York Times.

ABSTRACT:

Dynamic linear algebra — algorithmic techniques for matrices that change over time — lies at the core of many applications, from continuous optimization to efficient graph algorithms to machine learning. Yet, until recently, the full power of dynamic linear algebra was not known and exploited in most applications. 

In this talk, I will describe several new advances in using these techniques and outline the limits of what can be done with them. I will overview progress on longstanding problems in dynamic shortest path data structures, regression algorithms, optimal transport, and other problems using dynamic linear algebra.

BIO:

Jan van den Brand obtained his bachelor's and master's degrees in both mathematics and computer science at the Goethe University Frankfurt. Currently, he is a Ph.D. candidate at KTH Royal Institute of Technology, Stockholm, and recipient of the Google Ph.D. Fellowship. His research is on efficient algorithms with focus on optimization and dynamic problems.

WATCH HERE: https://bluejeans.com/558696336

ABSTRACT:

BIO:

ABSTRACT:

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 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 Ph.D. 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.

Watch here.

Registration is required. Register here.

Registration is required. Register here.

Scalable and Responsible AI to Find the Right Career for the Right Person

Abstract:

Eightfold's mission statement is to find the right career for the right person. In order to successfully find the right career for any individual, it is imperative to have a deep understanding of their past and current accomplishments and use them to forecast their future capabilities and potential. The Machine Learning at Eightfold tackles this exact problem that helps to hire for potential and answer the question "What’s Next for You"

Bio:

Sanjeet Hajarnis is a Principal Engineer at Eightfold AI and focuses on building scalable Machine Learning platforms at Eightfold. At Eightfold, he focuses on building large scale language models. Prior to Eightfold, Sanjeet has worked at Uber (Matching) and Facebook (News Feed). He graduated from Georgia Tech in 2012 with a specialization in Machine Learning.

The summit will start at 1:00pm – 5:00 p.m., Thursday, Jan. 28^, and will resume at 12:00 p.m. – 5:00 p.m. on Friday, Jan. 29.   

The summit will start at 1:00pm – 5:00 p.m., Thursday, Jan. 28^, and will resume at 12:00 p.m. – 5:00 p.m. on Friday, Jan. 29.   

ABSTRACT: 

Over the last two decades, software engineering (broadly construed to include testing, analysis, synthesis, verification, and security) has witnessed a silent revolution in the form of Boolean SAT and SMT solvers. These solvers are now integral to many testing, analysis, synthesis, and verification approaches. This is largely due to adramatic improvement in the scalability of these solvers vis-a-vis large real-world formulas. What is surprising is that the Booleansatisfiability problem is NP-complete, believed to be intractable, and yet these solvers easily solve industrial instances containing millions of variables and clauses in them. How can that be?

In my talk, I will address this question of why SAT solvers are soefficient through the lens of machine learning (ML) as well as ideas from (parameterized) proof complexity. While the focus of my talk is almost entirely empirical, I will show how can we leverage theoretical ideas to not only deepen our understanding but also to build better SAT solvers. I will argue that SAT solvers are best viewed as proof systems, composed of two kinds of sub-routines, ones that implement proof rules and others that are prediction engines that optimize some metric correlated with solver running time. These prediction engines can be built using ML techniques, whose aim is to structure solverproofs in an optimal way. Thus, two major paradigms of AI, namely ML and logical deduction, are brought together in a principled way in order to design efficient SAT solvers. A result of my research is the MapleSAT solver, that has been the winner of several recent international SAT competitions and is widely used inindustry and academia.

BIO:

Dr. Vijay Ganesh is an associate professor at the University of Waterloo. Prior to joining Waterloo in 2012, he was a research scientist at MIT (2007-2012) and completed his Ph.D. in computer science from Stanford in 2007. Ganesh’s primary area of research is the theory and practice of automated mathematical reasoning algorithms aimed at software engineering, security, and mathematics. In this context he has led the development of many SAT/SMT solvers, most notably, STP, Z3 string, MapleSAT, and MathCheck. He has also proved several decidability and complexity results in the context of first-order theories. He has won over 25 awards, honors, and medals to date for his research, including an ACM Impact Paper Award at ISSTA 2019, ACM Test of Time Award at CCS 2016, and a Ten-Year Most Influential Paper citation in 2008. He is the editor-in-chief of the Springer book series Progress in Computer Science and Applied Logic (PCSAL) and has co-chaired many conferences, workshops, and seminars including a Simons Institute semester on Boolean Satisfiability in 2021.

Please join us a virtual Dean's New Undergraduate Alumni Celebration, to be held Thursday, December 10 at 7 p.m. EST on BlueJeans**

If you are a graduate, be sure to look for an email titled "create your personal graduation slide" to personalize your slide for the virtual celebration! Wednesday, Dec. 2 at 5 p.m. EST is the slide submission deadline - there is NO flexibility in this submission deadline.

The program for December 10 is as follows;

• Charles Isbell, Dean of the College of Computing, John P. Imlay Jr. Chair
• Slideshow presentation of all the Spring 2020 graduates in alpha order by last name 
• Cedric Stallworth, Assistant Dean for Outreach, Enrollment and Community; Senior Lecturer
• Jennifer Whitlow, Director of Computing Enrollment and Alumni Engagement

**This ceremony will also be housed on our website and emailed to you after the event.

Be sure to tag your Dean’s New Alumni Celebration photos and posts on Instagram and Twitter with #GTComputing and #GT20

We wish all of our graduates a safe, healthy and enjoyable winter!

Please join us for a virtual Dean's New Graduate Alumni Celebration, to be held on Monday, December 14 at 7 p.m. EST on BlueJeans**

The program for December 14 is as follows;

• Charles Isbell, Dean of the College of Computing, John P. Imlay Jr. Chair
• Slideshow presentation of all the Fall 2020 graduates in alpha order by last name 
• David Joyner, Executive Director of Online Education & OMSCS and Senior Research Associate
• Jennifer Whitlow, Director of Computing Enrollment and Alumni Engagement

**This ceremony will also be housed on our website afterwards. 

If you are a graduate, be sure to look for an email titled "create your personal graduation slide" to personalize your slide for the virtual celebration! Wednesay, Dec. 2 at 5 p.m. EST is the slide submission deadline - there is NO flexibility in this submission deadline.

If you are an OMS graduate, please tune in HERE Monday, December 14 at 10 a.m. EST or HERE Monday, December 14 at 6 p.m. EST to see Executive Director, David Joyner give a tour of the Georgia Tech campus. You will have the opportunity to ask questions during this live event. Any and all are welcome to join. Just so we have an idea of numbers, please RSVP HERE. This tour will also be hosted on the website a week after the event.
Be sure to tag your Dean’s New Alumni Celebration photos and posts on Instagram and Twitter with #GTComputing and #GT20

We wish all of our graduates a safe, healthy, and enjoyable winter!

Register at: https://primetime.bluejeans.com/a2m/register/kgjkcdxt

Talk title:

Bringing Visual Memories to Life

Abstract:

Photography allows us to capture and share memorable moments of our lives. However, 2D images appear flat due to the lack of depth perception and may suffer from poor imaging conditions such as taking photos through reflecting or occluding elements. In this talk, I will present our recent efforts to overcome these limitations. Specifically, I will cover our recent for creating compelling 3D photography, removing unwanted obstructions seamlessly from images or videos, and estimating consistent video depth for advanced video-based visual effects. I will conclude the talk with some ongoing research and research challenges ahead. 

Speaker: 

- Jia-Bin Huang, Assistant Professor, Virginia Tech

- Website: https://filebox.ece.vt.edu/~jbhuang/

Bio:

Jia-Bin Huang is an Assistant Professor in the Bradley Electrical and Computer Engineering at Virginia Tech. He received his Ph.D. degree from the Department of Electrical and Computer Engineering at the University of Illinois, Urbana-Champaign. His research interests include computer vision, computer graphics, and machine learning with a focus on visual analysis and synthesis with physically grounded constraints. His research received the best student paper award in IAPR International Conference on Pattern Recognition (ICPR) for the work on computational modeling of visual saliency and the best paper award in the ACM Symposium on Eye Tracking Research & Applications (ETRA) for work on learning-based eye gaze tracking. Huang is the recipient of the NSF CRII award, Samsung Global Outreach Award, 3M non-tenured faculty award, and a Google faculty research award. 

Registration is required.

Participating labs include:

• Yao’s Group - Yao Xie, H. Milton Stewart School of Industrial Systems and Engineering (ISyE)
• Huo Lab - Xiaoming Huo, ISyE
• FLASH – Tuo Zhao, ISyE and School of Computational Science and Engineering (CSE)
• LF Radio Lab – Morris Cohen, School of Electrical Computing and Engineering (ECE)
• Polo Club of Data Science – Polo Chau, CSE
• Network Science – Constantine Dovrolis, School of Computer Science
• CLAWS – Srijan Kumar, CSE
• Rehg Lab – Jim Rehg, School of Interactive Computing (IC)
• Control, Optimization, Algorithms, and Randomness (COAR) Lab – Siva Theja Maguluri, ISyE
• Entertainment Intelligence Lab and Human Centered AI Lab – Mark Riedl, IC
• Social and Language Technologies (SALT) Lab – Diyi Yang, IC
• FATHOM Research Group – Swati Gupta, ISyE
• Zhang's CompBio Lab – Xiuwei Zhang, CSE
• Statistical Machine Learning - Ashwin Pananjady, ISyE and ECE
• AdityaLab - B. Aditya Prakash, CSE
• OLIVES - Ghassan AlRegib, ECE
• Robotics Perception and Learning (RIPL) – Zsolt Kira, IC
• Eye-Team - Irfan Essa, IC
• Mark Davenport, ECE

Time: 3:00pm-4:00pm Eastern Time (ET)

Join virtually: https://primetime.bluejeans.com/a2m/live-event/xvrbevrh

Title

LinkedIn Data Science - Creating Global Economic Opportunities

Abstract

At LinkedIn, data plays an essential role in achieving our vision of creating economic opportunity for every member of the global workforce. In this talk, Ya Xu will share perspectives from her experience and will highlight a few interesting problems her team tackles, such as measuring network effects, addressing cannibalization bias on advertising using budget split, forecasting, data privacy, responsible design, and fairness in our products and ML models.

Speaker

Ya Xu, VP of Engineering and Head of Data Science Team at LinkedIn

Ya joined LinkedIn in 2013 and has since truly helped make LinkedIn a Data First company. She leads an exceptional team of talented data scientists whose work covers metrics, insights, inference and algorithms and they tackle data science challenges across product, sales, marketing, economics, infrastructure, and operations. This centralized group has 300+ data scientists distributed across US (Sunnyvale, Mountain View, San Francisco, New York), India, China, Singapore and Dublin, Ireland. Ya is passionate about bridging science and engineering to create impactful results. She and her team try to keep LinkedIn on the cutting edge while ensuring that its A.I. systems avoid providing biased results while maintaining user privacy. They help the company take active responsibility over the data they collect to ensure fairness and protect privacy.

In addition to her work at LinkedIn, Ya’s contributions outside of her day job, such as the bookshe co-authored on Experimentation and her Stanford commencement speech, are meaningful to the entire industry as well as future Data Scientists. Before LinkedIn, she worked at Microsoft and received a PhD in Statistics from Stanford University.

Ya Xu’s LinkedIn Profile: https://www.linkedin.com/in/ya-xu/

Host

Polo Chau

Associate Professor, CSE, College of Computing

Associate Director, MS Analytics

Director of Industry Relations, Institute for Data Engineering and Science (IDEaS)

Associate Director of Corporate Relations, Center for Machine Learning (ML@GT)

2018-2021 Provost Teaching and Learning Fellow

The panel will feature thought leaders from Google, Georgia Tech, and Queer in AI, who will together answer questions like "What implications and problems exist or will exist if the tech workforce does not become more diverse?" and "How does anyone make sure they are not introducing their bias into a given system? What questions should we be asking or actions should we be taking to avoid this?"

Registration is required and available here.

Panelists

Charles Isbell

Charles Isbell is the Dean of Computing and The John P. Imlay Jr. Chair at Georgia Tech’s College of Computing. He is also a professor in the School of Interactive Computing and the Machine Learning Center at Georgia Tech (ML@GT.)

Isbell's research passion is artificial intelligence. In particular, he focuses on applying statistical machine learning to building autonomous agents that must live and interact with large numbers of other intelligent agents, some of whom may be human.

Lately, Isbell has turned his energies toward adaptive modeling, especially activity discovery (as distinct from activity recognition); scalable coordination; and development environments that support the rapid prototyping of adaptive agents. As a result, he has begun developing adaptive programming languages, worrying about issues of software engineering, and trying to understand what it means to bring machine learning tools to non-expert authors, designers, and developers.

Rapha Gontijo Lopes

Rapha Gontijo Lopes is a research associate at Google Brain and a founder of Queer in AI. He joined Google as an AI Resident in the summer of 2018 after completing his B.S. degree in Computer Science at Georgia Tech. His work investigates how deep learning works (or doesn't), and the role that input distributions have on model robustness. As an organizer for Queer in AI, he creates academic workshops that drive the research conversation at the intersection of AI and LGBTQ+ people.

Tiffany Deng

Tiffany Deng leads the Responsible AI Program Management Team at Google where she is focused on helping people build products that work for everyone. Prior to Google, Tiffany worked as a Privacy Program Manager at Facebook and as consultant in Washington, D.C.

Moderator: Deven Desai

Deven Desai is an associate director of the Machine Learning Center (ML@GT) and associate professor in the Scheller College of Business. He was also the first, and to date, only Academic Research Counsel at Google, Inc., and a Visiting Fellow at Princeton University's Center for Information Technology Policy.

 Professor Desai's scholarship examines how business interests, new technology, and economic theories shape privacy and intellectual property law and where those arguments explain productivity or where they fail to capture society's interest in the free flow of information and development

REGISTER HERE

Title: Towards High Precision Text Generation

Abstract:

Despite large advances in neural text generation in terms of fluency, existing generation techniques are prone to hallucination and often produce output that is unfaithful or irrelevant to the source text. In this talk, we take a multi-faceted approach to this problem from 3 aspects: data, evaluation, and modeling. 

From the data standpoint, we propose ToTTo, a tables-to-text-dataset with high quality annotator revised references that we hope can serve as a benchmark for high precision text generation task.  While the dataset is challenging, existing n-gram based evaluation metrics are often insufficient to detect hallucinations. To this end, we propose BLEURT, a fully learnt end-to-end metric based on transfer learning that can quickly adapt to measure specific evaluation criteria. Finally, we propose a model based on confidence decoding to mitigate hallucinations. 

Collaborators: This is joint work with Thibault Sellam, Ran Tian, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 

About the author:

Ankur Parikh is a senior research scientist at Google NYC and adjunct assistant professor at NYU. His research interests are in natural language processing and machine learning with a recent focus on high precision text generation. Ankur received his PhD from Carnegie Mellon in 2015 and has received a best paper runner up award at EMNLP 2014 and a best paper in translational bioinformatics at ISMB 2011.

Register Here

Reasoning about Complex Media from Weak Multi-modal Supervision

Abstract:

In a world of abundant information targeting multiple senses, and increasingly powerful media, we need new mechanisms to model content. Techniques for representing individual channels, such as visual data or textual data, have greatly improved, and some techniques exist to model the relationship between channels that are “mirror images” of each other and contain the same semantics. However, multimodal data in the real world contains little redundancy; the visual and textual channels complement each other. We examine the relationship between multiple channels in complex media, in two domains, advertisements and political articles.

First, we collect a large dataset of advertisements and public service announcements, covering almost forty topics (ranging from automobiles and clothing, to health and domestic violence). We pose decoding the ads as automatically answering the questions “What should do viewer do, according to the ad” (the suggested action), and “Why should the viewer do the suggested action, according to the ad” (the suggested reason). We train a variety of algorithms to choose the appropriate action-reason statement, given the ad image and potentially a slogan embedded in it. The task is challenging because of the great diversity in how different users annotate an ad, even if they draw similar conclusions. One approach mines information from external knowledge bases, but there is a plethora of information that can be retrieved yet is not relevant. We show how to automatically transform the training data in order to focus our approach’s attention to relevant facts, without relevance annotations for training. We also present an approach for learning to recognize new concepts given supervision only in the form of noisy captions.

Second, we collect a dataset of multimodal political articles containing lengthy text and a small number of images. We learn to predict the political bias of the article, as well as perform cross-modal retrieval despite large visual variability for the same topic. To infer political bias, we use generative modeling to show how the face of the same politician appears differently at each end of the political spectrum. To understand how image and text contribute to persuasion and bias, we learn to retrieve sentences for a given image, and vice versa. The task is challenging because unlike image-text in captioning, the images and text in political articles overlap in only a very abstract sense. We impose a loss requiring images that correspond to similar text to live closeby in a projection space, even if they appear very diverse purely visually. We show that our loss significantly improves performance in conjunction with a variety of existing recent losses. We also propose new weighting mechanisms to prioritize abstract image-text relationships during training.

Bio:

Adriana Kovashka is an Assistant Professor in Computer Science at the University of Pittsburgh. Her research interests are in computer vision and machine learning. She has authored eighteen publications in top-tier computer vision and artificial intelligence conferences and journals (CVPR, ICCV, ECCV, NeurIPS, AAAI, ACL, TPAMI, IJCV) and ten second-tier conference publications (BMVC, ACCV, WACV). She has served as an Area Chair for CVPR in 2018-2021, NeurIPS 2020, ICLR 2021, AAAI 2021, and will serve as co-Program Chair of ICCV 2025. She has been on program committees for over twenty conferences and journals, and has co-organized seven workshops. Her research is funded by the National Science Foundation, Google, Amazon and Adobe.

Abstract

The healthcare industry is rapidly evolving and this change is fueled by data. Now more than ever, data management ensures data-driven personalization for each individual patient. Learn how Anthem uses data to improve the lives of each of our patients. During this session, our team will discuss how data is ingested, processed, cleansed, managed, and visualized to improve healthcare for our communities.

Speakers

Sanjay Vishwakarma, Sr. Director Engineering

Sanjay Vishwakarma is the Engineering Senior Director for Data and Analytics Operations within Anthem. In this role, Sanjay architects complex solutions for Anthem’s Data Lake including: data ingestion, standardization, curation, and consumption in both onPrem and Cloud based implementations. Prior to joining Anthem, Sanjay served in enterprise architecture roles spanning several industries including healthcare, insurance, technology, and the financial industry.

Shannon Miles, Director

Shannon Miles has over 23 years of data warehousing leadership with Anthem. Shannon has owned large-scale initiatives across all lines of business and multiple domains. In his current role, Shannon is the Director within the IT Program Integrity team. In this role, Shannon leads program integrity efforts utilizing multiple platforms including analytics based claims and provider analysis.

Host 

Polo Chau

Associate Professor, CSE, College of Computing

Associate Director, MS Analytics

Director of Industry Relations, Institute for Data Engineering and Science (IDEaS)

Associate Director of Corporate Relations, Center for Machine Learning (ML@GT)

polo@gatech.edu

http://www.cc.gatech.edu/~dchau/

Abstract

The healthcare industry is rapidly evolving and this change is fueled by data. Now more than ever, data management ensures data-driven personalization for each individual patient. Learn how Anthem uses data to improve the lives of each of our patients. During this session, our team will discuss how data is ingested, processed, cleansed, managed, and visualized to improve healthcare for our communities.

Speakers

Sanjay Vishwakarma, Sr. Director Engineering

Sanjay Vishwakarma is the Engineering Senior Director for Data and Analytics Operations within Anthem. In this role, Sanjay architects complex solutions for Anthem’s Data Lake including: data ingestion, standardization, curation, and consumption in both onPrem and Cloud based implementations. Prior to joining Anthem, Sanjay served in enterprise architecture roles spanning several industries including healthcare, insurance, technology, and the financial industry.

Shannon Miles, Director

Shannon Miles has over 23 years of data warehousing leadership with Anthem. Shannon has owned large-scale initiatives across all lines of business and multiple domains. In his current role, Shannon is the Director within the IT Program Integrity team. In this role, Shannon leads program integrity efforts utilizing multiple platforms including analytics based claims and provider analysis.

Host 

Polo Chau

Associate Professor, CSE, College of Computing

Associate Director, MS Analytics

Director of Industry Relations, Institute for Data Engineering and Science (IDEaS)

Associate Director of Corporate Relations, Center for Machine Learning (ML@GT)

polo@gatech.edu

http://www.cc.gatech.edu/~dchau/

Register: https://primetime.bluejeans.com/a2m/register/ppyhwquh

This event will take place on Friday, October 30 at 2 pm ET// 11 am PT.

Emerging technologies, particularly within computer vision, photogrammetry, and spatial computing, are unlocking new forms of storytelling for journalists to help people understand the world around them. In this talk, members of the R&D team at The New York Times talk about their process for researching and developing new capabilities built atop emerging research. In particular, hear how they are embracing photogrammetry and spatial computing to create new storytelling techniques that allow a reader to experience an event as close to reality as possible. Learn about the process of collecting photos, generating 3D models, editing, and technologies used to scale up to millions of readers. The team will also share their vision for these technologies and journalism, their ethical considerations along the way, and a research wishlist that would accelerate their work. 

In its 169 year history, The New York Times has evolved with new technologies, publishing its first photo in 1896 with the rise of cameras, introducing the world’s first computerized news retrieval system in 1972 with the rise of the computer, and launching a website in 1996 with the rise of the internet. Since then, the pace of innovation has accelerated alongside the rise of smartphones, cellular networks, and other new technologies. The Times now has the world’s most popular daily podcast, a new weekly video series, and award-winning interactive graphics storytelling. Join us for a discussion about how our embrace of emerging technologies is helping us push the boundaries of journalism in 2020 and beyond.

Speakers:

Members of The New York Times R&D Team, including:

• Marc Lavallee, Executive Director, R&D

• Or Fleisher, Sr. Engineer

• Mint Boonyapanachoti, Creative Technologist

• Lana Porter, Creative Director

• Mark McKeague, Technical Lead

Team Bio:

The New York Times Research and Development team applies emerging technologies in service of our company’s mission to seek the truth and help people understand the world.

Using new technologies and formats, we develop technical capabilities for our newsroom and new forms of storytelling for our readers. As part of our method, we evaluate emerging trends in media and technology and forecast how they might play out over the next two to three years. Once we identify an opportunity, we dedicate a team to explore the space and develop products in collaboration with other parts of our organization.

We're a multidisciplinary team of journalists, creative technologists, designers, and engineers.

Robert Nowak will give a virtual seminar on October 7, 2020. Please check back soon for registration and talk details.

Register: https://primetime.bluejeans.com/a2m/register/vzjxvxjh

Abstract:

The field of Machine Learning (ML) has advanced considerably in recent years, but mostly in well-defined domains using huge amounts of human-labeled training data. Machines can recognize objects in images and translate text, but they must be trained with more images and text than a person can see in nearly a lifetime.  The computational complexity of training has been offset by recent technological advances, but the cost of training data is measured in terms of the human effort in labeling data. People are not getting faster nor cheaper, so generating labeled training datasets has become a major bottleneck in ML pipelines. 

Active ML aims to address this issue by designing learning algorithms that automatically and adaptively select the most informative examples for labeling so that human time is not wasted labeling irrelevant, redundant, or trivial examples. This talk explores the development of active ML theory and methods over the past decade, including a new approach applicable to kernel methods and neural networks, which views the learning problem through the lens of representer theorems. This perspective highlights the effect that adding a given training example has on the representation.   The new approach is shown to possess a variety of desirable mathematical properties that allow active learning algorithms to learn good classifiers from few labeled examples.

About Robert:

Nowak holds the Nosbusch Professorship in Engineering at the University of Wisconsin-Madison, where his research focuses on signal processing, machine learning, optimization, and statistics.

Through the looking glass: what NLP can reveal about sociolinguistic variation

Abstract:

People adapt their language use in different social contexts to meet communicative needs: a person may use the word <going> with colleagues and <goin'> with their close friends. Sociolinguistics researchers investigate the systematic variation in language use across different contexts to determine the social meaning of variation, such as how people change their word choices for different audiences. While traditional sociolinguistics investigates variation in spoken language, computational sociolinguistics relies on natural language processing and statistical methods to investigate written language in online discussions. This talk will explore how NLP can help isolate sociolinguistic phenomena that would otherwise go understudied in spoken contexts, and more broadly how NLP can help social science research.

Bio:

Ian Stewart is a Visiting Research Investigator with the LIT Lab at the University of Michigan. Ian recently defended his thesis in the Human-Centered Computing PhD program at the Georgia Institute of Technology, and his research proposed computational approaches to understanding sociolinguistic variation on social media. Ian is currently interested in incorporating sociolinguistic insight into NLP models to address different speaker needs.

Register at: https://primetime.bluejeans.com/a2m/register/rujhjqhe

Title:

Machine Learning Challenges at Google Ads

Abstract:

While we have made significant advances in the last two decades, serving the most relevant ads to our users is still a big challenge to date. In particular, commercial contents are full of images, videos in addition to textual information, and understanding our advertiser products and offers from their ad creatives and landing pages poses interesting multimodal modeling problems. In this talk, we will discuss a new multimodal task (Multi-Image Summarization) and the associated dataset we are releasing. We will also talk about our recent work which co-embeds text and image in a shared embedding space to improve a cross-modal retrieval task. Then we will share some challenges & experiences from quality improvements in Search Ads products.

About Pradyumna

Pradyumna Narayana is a software engineer for Google. He joined Google in 2018 and is conducting research at the intersection of computer vision and natural language processing for Search Ads. He earned his PhD in the area of Computer Vision from Colorado State University in 2018. His Ph.D work focused on Gesture recognition from videos using Deep Learning. Prior to that, he earned his MS from Colorado State University in 2015.

About Kazoo

Kazoo Sone is a software engineer for Google. Since he joined Google in 2011, he has led several machine learning & natural language processing projects for Search Ads and Research and contributed to many of Google’s Ads products. He earned his MS from Georgia Tech and Ph.D from Caltech.

Register: https://primetime.bluejeans.com/a2m/register/ghfqxrfr

Title

Using rationales and influential training examples to (attempt to) explain neural predictions in NLP 

Abstract

Modern deep learning models for natural language processing (NLP) achieve state-of-the-art predictive performance but are notoriously opaque. I will discuss recent work looking to address this limitation. I will focus specifically on approaches to: (i) Providing snippets of text (sometimes called "rationales") that support predictions, and; (ii) Identifying examples from the training data that influenced a given model output. 

About Byron

Byron Wallace is an assistant professor in the Khoury College of Computer Sciences at Northeastern University. He earned his PhD from Tufts University in 2012, after which he taught at Brown University as research faculty. He joined Northeastern from the University of Texas at Austin, where he was an assistant professor in the School of Information from 2014-2016.

Wallace’s research areas include artificial intelligence, data science, machine learning, natural language processing, and information retrieval, with emphasis on applications in health informatics. Byron is a member of the applied machine learning group and the Data Science and Analytics Lab at Northeastern.

Wallace develops machine learning and natural language processing methods that make synthesizing the vast biomedical evidence-base more efficient. He also works on core machine learning and natural language processing methods, with his more of his recent work concerning Convolutional Neural Network (CNN) architectures for text. Wallace has recently been developing hybrid, interactive human/machine learning systems that aim to robustly combine human and machine intelligence.

His work has been supported by grants from the Army Research Office, the NIH, and the NSF. He won the Tufts University 2012 Outstanding Graduate Researcher award and his thesis work was recognized as The Runner Up for the 2013 ACM Special Interest Group on Knowledge Discovery and Data Mining (SIG KDD) Dissertation Award. He recently co-authored the winning submission for the Health Care Data Analytics Challenge at the 2015 IEEE International Conference on Healthcare Informatics.

The event will virtually take place via Bluejeans Events and requires registration.

About the Panelists:

Nicoleta Serban is the Virginia C. and Joseph C. Mello Professor in the H. Milton Stewart School of Industrial and Systems Engineering at Georgia Tech.

Dr. Serban's most recent research focuses on model-based data mining for functional data, spatio-temporal data with applications to industrial economics with a focus on service distribution and nonparametric statistical methods motivated by recent applications from proteomics and genomics. 

She received her B.S. in Mathematics and an M.S. in Theoretical Statistics and Stochastic Processes from the University of Bucharest. She went on to earn her Ph.D. in Statistics at Carnegie Mellon University. (Nicoleta's Work)

Aditya Prakash recently joined Georgia Tech as an associate professor in the School of Computational Science and Engineering. He has published one book, more than 80 papers in major venues, holds two U.S. patents and has given four tutorials at leading conferences. His work has received a best paper award and four best-of-conference selections. Tools developed by his group have been in use in many places including ORNL, Walmart and Facebook. His research interests include Data Science, Machine Learning and AI, with emphasis on big-data problems in large real-world networks and time-series. (Aditya's Work)

Munmun DeChoudhury is an associate professor in the School of Interactive Computing. She is affiliated with ML@GT, GVU Center, and IPaT. At Georgia Tech, she leads the Social Dynamics and Wellbeing Lab to study, analyze, and appropriate social media, responsibly and ethically to derive computational, large-scale data-driven insights, and to develop mechanisms and technologies for improving our well-being, particularly our mental health. Her research has been motivated by how the availability of large-scale online social data, with the amalgamation of advances in machine learning and grounding in human-centered approaches can help us answer fundamental questions relating to our social lives. (Munmun's Work)

Srijan Kumar is an assistant professor in the School of Computational Science and Engineering. His research develops data science solutions to address the high-stakes challenges on the web and in the society. He has pioneered the development of user models and network science tools to enhance the well-being and safety of people. His research has been the subject of a documentary and has been recognized with best paper awards at WWW and ICDM. (Srijan's Work) (Srijan's Work Part 2)

Kenneth Miller is a student in Georgia Tech’s Online Master’s of Computer Science (OMSCS) program. He is a partner at Erskine Law where he represents Ford Motor Company and cases in the field of mass toxic tort litigation. Miller is a 13-year veteran of the United States Navy. (Kenneth's Work)

Irfan Essa is the executive director of ML@GT and distinguished professor and senior associate dean in the School of Interactive Computing. Essa is also a senior staff research scientist at Google. His research focuses on computer vision, machine learning, computer graphics, computational perception, robotics, computer animation, and social computing. Essa is a IEEE Fellow and has published over 200 scholarly articles with several winning best paper awards.

Please join us for a virtual Dean's New Graduate Alumni Celebration, to be held Tuesday, May 19 at 2 p.m. EST on BlueJeans**

If you are a student, please create your profile no later than Tuesday, May 5 by 12 p.m. EST on the Marching Order website to be included in the slideshow during the program. If not, no profile needs to be created, just simply register and tune into the BlueJeans link or refer to our website at a later date.

**This ceremony will be live streamed HERE and will also be housed on our website a week after the ceremony. If you use closed captioning, there will be a link for this when you enter the live event on BlueJeans. 
 

ABSTRACT:

Machine learning plays a key role in shaping the decisions made by a growing number of institutions. This talk will share variational perspectives on aspects of inference, algorithms and fairness. On the topic of algorithms, I will present a variational framework on a classical family of convex optimization algorithm called accelerated gradient algorithms and demonstrate how it leads to simpler, faster gradient-based algorithms and generalizations of existing acceleration frameworks. On the topic of inference, I will present a variational framework for developing computationally eÿcient approximations of cross-validation and show how it provides fast and reliable estimates of out-of-sample performance for many machine learning models. On the topic of fairness, I will present a variational model for reasoning about the long-term impacts of using machine learning models to allocate scarce resources and opportunities to people, such as employment and educational decisions.

BIO:

Ashia Wilson is a postdoctoral researcher in the Machine Learning Group at Microsoft Research, New England. She received undergraduate degrees in applied mathematics and philosophy from Harvard University in 2011. She received her doctorate in atatistics from the University of California, Berkeley in 2018 advised by Benjamin Recht and Michael I. Jordan. Her research interests are in providing rigorous guarantees for algorithmic performance and developing frameworks for studying issues of fairness and governance in machine learning.

ABSTRACT:

This talk will discuss statistical and computation requirements — and how they interact — for three learning setups. In the first part, we inspect the role of memory in learning. We study how the total memory available to a learning algorithm affects the amount of data needed for learning (or optimization), beginning by considering the fundamental problem of linear regression. Next, we examine the role of long-term memory vs. short-term memory for the task of predicting the next observation in a sequence given the past observations. Finally, we explore the statistical requirements for the task of manufacturing more data — namely how to generate a larger set of samples from an unknown distribution. Can “amplifying” a dataset be easier than learning?

BIO:

Vatsal Sharan is a Ph.D. student at Stanford, advised by Greg Valiant. He is a part of the Theory group and the Statistical Machine Learning group, and his primary interests are in the theory and practice of machine learning.

ABSTRACT:

Reconfigurable analog devices are a powerful new computing substrate especially appropriate for executing dynamical systems in an energy efficient manner. These devices leverage the physical behavior of transistors to directly implement computation. Under this paradigm, voltages and currents within the device implement continuously evolving variables in the computation.

In this talk, I discuss compilation techniques for automatically configuring such devices to execute dynamical systems. I present Legno, the first compilation system that automatically targets a real reconfigurable analog device of this class. Legno synthesizes analog circuits from parametric and specialized analog blocks and accounts for analog noise, quantization error, operating range limitations, and manufacturing variations within the device. I evaluate Legno on applications from the biology, physics, and controls domains. The results demonstrate that these applications execute with acceptable error while consuming microjoules of energy.

BIO:

Sara Achour is a Ph.D. candidate at the Computer Science and Artificial Intelligence Laboratory at Massachusetts Institute of Technology (CSAIL MIT) and a NSF Fellowship recipient. Her research focuses on new techniques and tools, specifically new programming languages, compilers, and runtime systems that enable end users to easily develop computations that exploit the potential of emerging nontraditional computing platforms.