CS Department Colloquium Series

Technology scaling of DRAM cells has enabled higher capacity memory for the last few decades. Unfortunately, DRAM cells become vulnerable to failure as they scale down to a smaller size. Enabling high-performance, energy-efficient, scalable memory systems without sacrificing the reliability is a major research challenge. My work focuses on designing a scalable memory system by rethinking the traditional assumptions in abstraction and separation of responsibilities across system layers. 

In this talk, I will discuss three fundamental ways to enable DRAM scaling. First, we can enable scaling by letting the manufacturers build smaller cells without providing any strict reliability guarantee.  I envision manufacturers shipping DRAMs without fully ensuring correct operation, and the system being responsible for detecting and mitigating DRAM failures while operating in the field. However, designing such a system is difficult due to intermittent DRAM failures. In this talk, I will discuss a system design, capable of providing reliability guarantees even in the presence of intermittent failures. Second, we can enable high-capacity memory leveraging the emerging non-volatile memory technologies that are predicted to be more scalable. I will present my vision to redefine the hardware and operating system interface to unify memory and storage system with non-volatile memory and discuss the opportunities and challenges of such a system. Third, tolerating failures in the application can improve memory scalability. The fundamental challenge of such a system is how to assure, verify, and quantify the quality of the results. I envision a system that limits the impact of memory failures such that it is possible to statically determine the worst-case results from the maximum possible error in the input. 

Bio:
Samira Khan is an Assistant Professor at the University of Virginia (UVa). Prior to joining UVa, she was a Post Doctoral Researcher at Carnegie Mellon University, funded by Intel Labs. Her research focuses on improving the performance, efficiency, and reliability of the memory system. She is the recipient of NSF CRII Award, NSF GOALI Award, and Rising Stars in EECS Award. She received her PhD from the University of Texas at San Antonio. During her graduate studies, she worked at Intel, AMD, and EPFL. 

Computer systems are highly vulnerable; attackers everyday discover new security vulnerabilities and exploit them to compromise the target systems. This talk will present our approaches to automatically prevent software vulnerabilities from exploitation. In particular, this talk will describe in detail two classes of vulnerabilities: an emerging class, called "type confusion" (or "bad casting"), that are commonly seen in modern web browsers, and a new class that we discovered, called "uninitialized padding," causing information leakage in the Linux kernel. This talk will explain what these vulnerabilities are, how attackers exploit them, why/how developers introduced them, and why it is non-trivial to avoid them in complex, real-world programs. Finally, our approaches to automatically eliminate them in practice will be demonstrated.

Bio:
Taesoo Kim is a Catherine M. and James E. Allchin Early Career Assistant Professor in the School Computer Science at the Georgia Institute of Technology (Georgia Tech). He also serves as the director of the Georgia Tech Systems Software and Security Center (GTS3). He is genuinely interested in building a system that prioritizes security principles first and foremost. Those principles include the total design of the system, analysis of its implementation, elimination of certain classes of vulnerabilities, and clear separation of its trusted components.  His thesis work, in particular, focused on detecting and recovering from attacks on computer systems, known as "undo computing." Taesoo holds a S.M. (2011) and a Ph.D. (2014) from MIT.

Natural language is built from a library of concepts and compositional operators that provide a rich source of information about how humans understand the world.  Can this information help us build better machine learning models? In this talk, we'll explore three ways of integrating compositional linguistic structure and learning: using language as a source of modular reasoning operators for question answering, as a scaffold for fast and generalizable reinforcement learning, and as a tool for understanding representations in neural networks.

Bio:
Jacob Andreas is a fifth-year PhD student at UC Berkeley working in natural language processing. He holds a B.S. from Columbia and an M.Phil. from Cambridge, where he studied as a Churchill scholar. His papers were recognized at NAACL 2016 and ICML 2017. Jacob has been an NSF graduate fellow, a Huawei--Berkeley AI fellow, and a Facebook fellow.

Despite decades of research building secure operating systems, many deployed systems must still choose between flexible application APIs and security. As a result, the vast majority of programmers are unable to improve these systems. This is not merely a result of poor system building. It is hard to design highly  extensible systems that are both secure and useful. Moreover, evaluating novel  designs with actual developers is critical in order to make sure system builders  can adopt research systems in practice.

Fortunately, in emerging application domains, such as the Internet of Things, there are no entrenched operating systems and application portability is less important. This makes it possible evaluate research techniques for building more secure and extensible systems with developers who are willing to adopt them.

I'll describe Tock, an operating system for microcontrollers that enables third-party developers to extend the system. Tock uses the Rust type-system to isolate kernel extensions and the hardware to isolate applications. I'll discuss  how we continuously evaluate Tock by engaging with practitioners, and how  lessons from practitioners have fed back into the system's design.

Bio:
Amit Levy is a PhD candidate in Computer Science at Stanford University.  He builds secure operating system kernels, web platforms, and network systems  that help make computers more programmable by third-party application  developers.

How will self-driving cars change urban mobility? This talk describes contributions in machine learning and optimization critical for enabling mixed-autonomy mobility, the gradual and complex integration of automated vehicles into the existing transportation system. The talk first explores and quantifies the potential impact of a small fraction of automated vehicles on low-level traffic flow dynamics such as traffic congestion, using novel techniques in model-free deep reinforcement learning. Second, the talk presents generic reinforcement learning techniques for improved variance reduction, developed for large-scale control systems such as traffic networks and robotic manipulation. To anchor this work in a broader mobility planning context, the coordination of automated vehicles relies on accurate traffic flow sensing. To this end, a new convex optimization method for cellular network measurements from AT&T for all of California is introduced to address a flow estimation problem previously believed to be intractable. Finally, automated vehicles are expected to increase transportation demand through a phenomenon called induced demand. To address this, joint work with Microsoft Research is presented, which provides theoretical justification for the application of widely used clustering algorithms to ridesharing problems, designed to mitigate the strain on existing infrastructure. Together, these contributions demonstrate, through principled learning and optimization methods, that a small number of vehicles and sensors can be harnessed for significant impact on urban mobility.

Bio:
Cathy Wu is a PhD candidate in machine learning in Electrical Engineering and Computer Sciences (EECS) at UC Berkeley, the Berkeley Artificial Intelligence Research lab, Berkeley DeepDrive, California PATH, and the Berkeley RISELab. She is interested in developing principled computational tools to enable reliable and complex decision-making for critical societal infrastructure, such as transportation systems. Cathy received her Masters and Bachelors degrees in EECS from MIT.  She is the recipient of several fellowships including the NSF graduate fellowship, the Berkeley Chancellor's fellowship, the NDSEG fellowship, and the Dwight David Eisenhower graduate fellowship.  Her work was acknowledged by several awards, including the 2016 IEEE ITSC Best Paper Award and the 2017 ITS Outstanding Graduate Student Award. Her leadership, in particular as the Research Lead of the Learning Traffic Team at Berkeley, was recognized by numerous awards and invitations, such as the 2017 IEEE Leaders Summit and multiple NSF early-career investigator workshops on cyber-physical systems.  Throughout her career, Cathy has collaborated or interned broadly across fields, including civil engineering, mechanical engineering, urban planning, and public policy, and institutions, including OpenAI, Microsoft Research, the Google Self-Driving Car Team, AT&T, Facebook, and Dropbox.  As the founder and Chair of the Interdisciplinary Research Initiative within the ACM Future of Computing Academy, she is actively working to create international programs to further enable and support interdisciplinary research in computing.

AI has made huge advancement into our daily life and increasingly we require intelligent agents that work intimately with people in a changing environment. However, current systems mostly work in a passive mode: waiting for requests from users and processing them one at a time. An interactive agent must handle real-time, sequential inputs and actively collaborate with people through communication. In this talk, I will present my recent work addressing challenges in real-time language processing and collaborative dialogue. The first part involves making predictions with incremental inputs. I will focus on the application of simultaneous machine interpretation and show how we can produce both accurate and prompt translations. Then, I will present my work on building agents that collaborate with people through goal-oriented conversation. I will conclude by discussing future directions towards adaptive, active agents.

Bio:
He He is a postdoctoral researcher at Stanford University. She earned her Ph.D. in Computer Science at the University of Maryland, College Park. She is interested in natural language processing and machine learning. Her research focuses on building intelligent agents that work in a changing environment and interact with people, with an emphasis on language-related problems. Specific applications include dependency parsing, simultaneous machine interpretation, and goal-oriented dialogue. She is the recipient of the 2016 Larry S. Davis doctoral dissertation award.

Computing devices play a more significant role in our lives than ever before, so it is more important than ever for them to understand who we are as people on a deep level. But many of today’s user interfaces do not have such an understanding: rather, their designs are based on developers’ intuitions alone. This often leads to mismatches between how useful computing systems promise to be and how useful they are in practice.

In this talk, I will show how analyzing and even modeling human behavior can unlock insights that help resolve such mismatches, resulting in systems that are significantly more useful than what they would otherwise be. I will discuss four results that I have worked on: (1) a system for interacting with objects by looking at them, (2) a system for typing on smartphones more quickly and with much fewer errors, (3) a system that can recognize players and recommend video game levels from controller inputs alone, and (4) a system that makes it possible for people who are blind to play the same types of racing games that sighted players can play with the same speed and sense of control that sighted players have.

Bio:
Brian A. Smith is a Ph.D. candidate in Computer Science at Columbia University, where he is a member of the Computer Vision Laboratory and the Computer Graphics and User Interfaces Laboratory. His research is in human–computer interaction, accessibility, and game design, and focuses on analyzing human behavior to make computing more useful. He has been awarded the NDSEG Fellowship, an NSF IGERT data science traineeship, and Columbia Engineering’s Extraordinary Teaching Assistant Award. He received his MS and BS in Computer Science from Columbia University.

We are at an exciting point in the evolution of memory technology. Device manufacturers have created a new non-volatile memory (NVM) technology that can serve as both system memory and storage. NVM supports fast reads and writes similar to volatile memory, but all writes to it are persistent like a solid-state disk. The advent of NVM invalidates decades of design decisions that are deeply embedded in today's database management systems (DBMSs). These systems are unable to take full advantage of NVM because their internal architectures are predicated on the assumption that memory is volatile. With NVM, many of the components of today's DBMSs are unnecessary and will degrade the performance of data-intensive applications. Thus, the best way to resolve these shortcomings is by designing a new system explicitly tailored for NVM.

In this talk, I will present my research on the design and development of an NVM DBMS, called Peloton. Peloton's architecture shows that the impact of NVM spans across all the layers of the DBMS. I will first introduce write-behind logging, an NVM-centric protocol that improves the availability of the database system by two orders-of-magnitude compared to the widely-used write-ahead logging protocol. I will then present the BzTree, an NVM-centric index data structure that illustrates how to simplify programming on NVM. In drawing broader lessons from this work, I will argue that all types of software systems, including file systems, machine-learning systems, and key-value stores, are amenable to similar architectural changes to achieve high performance and availability on NVM.

Bio:
Joy Arulraj is a Ph.D. candidate at Carnegie Mellon University. His research interests are in database systems and data science. As part of his dissertation work, he has studied and built the first non-volatile memory database system, called Peloton, for performing large-scale transaction processing and real-time data analytics. He is the recipient of the Carnegie Mellon Presidential Fellowship and a Samsung Ph.D. Fellowship.

Concurrency, the processing of multiple requests simultaneously, is one of the most challenging problems large-scale server applications face in practice. Accordingly, database systems research has long made the case for automatically handling concurrency in the database by guaranteeing serializability. Serializability shields applications from reasoning about concurrency, and allows developers to focus entirely on implementing application logic. Unfortunately, in the 40-plus years since its inception, serializability has not seen wide adoption in practice. This is because weaker guarantees, which expose applications to concurrency and the inevitable bugs that arise, perform significantly better.
 
In this talk, I will discuss my research on addressing the performance limitations of serializability via deterministic transaction execution. Deterministic transaction execution exploits the fact that a large class of modern server applications do not require the full generality of conventional database transactions. By tailoring transaction execution mechanisms for this class of applications, my research shows that it is possible to achieve serializability with minimal performance overhead. I will first describe a serializable multi-versioning mechanism that decouples conflicting reads and writes, and subsequently outperforms a state-of-the-art implementation of the weaker guarantee of snapshot isolation by over 3x. Next, I will describe piecewise visibility, a concurrency control mechanism that isolates requests at a finer granularity than entire transactions, which consequently permits aggressive serializable transaction interleavings and outperforms the weaker guarantee of read committed by over 3x. Finally, I will discuss ongoing work that applies deterministic transaction execution principles to address replication lag in Facebook’s production MySQL infrastructure.
 
Bio: 
Jose Faleiro is a PhD candidate in computer science at Yale University. His research interests are in data management systems, multi-core systems, and distributed systems. His thesis research investigates the use of deterministic execution to enable scalable and efficient transaction processing on main-memory multi-core database systems. In addition to his academic research at Yale, Jose has worked on large-scale real world systems, including Microsoft’s Orleans cloud programming framework, and Facebook’s production MySQL infrastructure. He is the recipient of the Alan J. Perlis Fellowship at Yale, and a Microsoft Research Tech Transfer Award. He has an undergraduate degree in computer science from the Birla Institute of Technology and Science (BITS), Pilani, India.

In this talk I will present my research on making computers see and think, two abilities essential for deep understanding of pixels. In particular, I will focus on understanding not just visual objects but also their connections. First, I will describe how we built ImageNet, a large-scale visual knowledge base that led to a paradigm shift in computer vision research. Next, I will show the design of a reasoning engine that uses an object relation graph to perform probabilistic logical inference in visual recognition. Third, I will present a general and powerful method for extracting deep semantics from pixels, with state-of-the-art results on multiple challenging tasks. I will conclude my talk by describing ongoing efforts and future directions toward deeper semantics, richer knowledge, and more human-like reasoning.

Bio: 
Jia Deng is an Assistant Professor of Computer Science and Engineering at the University of Michigan. His research focus is on computer vision and machine learning. He received his Ph.D. from Princeton University and his B.Eng. from Tsinghua University, both in computer science. He is a recipient of the Sloan Research Fellowship, PAMI Mark Everingham Prize, Yahoo ACE Award, Google Faculty Research Award, Amazon Research Award, ICCV Marr Prize, and ECCV Best Paper Award.