Spring 2018


February 8, 2018

We present our work on the extraction and estimation of syntactic paraphrases using commodity text data and automated linguistic annotation. Our initial approach leverages bilingual parallel data and builds on SCFG extraction techniques used in machine translation. We then extend our estimation methods to include contextual similarity metrics drawn from vast amount of plain text. We evaluate the quality of our paraphrases by applying a generalizable adaptation scheme that tunes our paraphraser to arbitrary text-to-text generation tasks, produce competitive results with only little data and work needed. We further discuss the scaling of our extraction method to large data sizes, and the building of the paraphrase database PPDB. a large-scale collection of paraphrases in 23 languages.

Speaker Biography: Juri‘s work has transitioned from working on Language Modeling, to Machine Translation, to Paraphrasing, to Semantic Parsing. His person transitioned from Ukraine to Germany, France, and the U.S. He’ll likely keep transitioning in one way or another.

Advisor: Chris Callison-Burch

Video Recording >>

February 20, 2018

It is notoriously difficult to understand, troubleshoot, and enforce distributed systems behaviors, because unlike standalone programs, they lack a central point of visibility and control. This impacts a range of important distributed systems management tasks, such as resource management, performance, security, accounting, and many more. In this talk I will outline several techniques and abstractions that let us re-establish this missing visibility and control, and reason about systems end-to-end. To demonstrate, I will present two management tools that I have developed in my research: Retro, which measures resource usage and co-ordinates scheduler parameters to achieve end-to-end performance goals; and Pivot Tracing, which dynamically monitors and correlates metrics across component boundaries. Together, these tools illustrate some of the common challenges and potential solutions when developing and deploying tools for distributed systems.

Speaker Biography: Jonathan Mace is a Ph.D. candidate in the Computer Science department at Brown University. His research centers on how to understand and enforce end-to-end behaviors in distributed systems. During his Ph.D. he was awarded the Facebook Fellowship in Distributed Systems, and he received an SOSP Best Paper Award for his work on Pivot Tracing.

Video Recording >>

February 22, 2018

Failures in medical devices, banking software, and transportation systems have lead to both significant fiscal costs and even loss of life. Researchers have developed sophisticated methods to monitor and understand many of the complex system mis-behaviors behind these bugs, but their computational costs (often an order of magnitude or more) prohibit their use in production, leading to an ecosystem of critical software with little guaranteed protection, and no method of reconciling misbehaviors.

In this talk I present systems and techniques which reduce the run-time burden of the tools required to understand and monitor the complex behaviors of today’s critical systems. First, I present Optimistic Hybrid Analysis (OHA). OHA observes that when applying static analysis towards dynamic analysis optimization, the static analysis need not be correct in all cases, so long as any analysis errors can be caught at runtime. This observation enables the use of much more efficient and accurate static analyses than historically used, creating dynamic run-times dramatically lower than prior techniques. Second, I argue that computer systems should be capable of not only recalling any prior state, but also providing the provenance of any byte within the history of the computation. I call such a system an “Eidetic System”, and I present Arnold, the first practical eidetic system, capable of recording and recalling years of computation on a single disk. I show that Arnold can practically answer critical questions about serious information leakages, such as exactly what information (if any) was leaked by the Heartbleed vulnerability, or Equifax breach.

Speaker Biography: David Devecsery

David Devecsery is currently a postdoctoral researcher at the University of Michigan, after completing his Ph.D. in January 2018 at the University of Michigan. His interests broadly span the areas of software systems, program analysis, and system security. David is particularly interested in creating practical tools that enable developers, users, and system administrators to practically observe and understand complex and unexpected behaviors of software systems.

Video Recording >>

February 27, 2018

Mobile devices are now the most common way users handle digital information and interact with online services. Different actors, trusting each other in different ways, compose the mobile ecosystem. Users interact with apps, trusting them to access valuable and privacy-sensitive information. At the same time, apps usually communicate with remote backends and authenticate users to online services. Finally, all these interactions are mediated, on one side, by the user interface and, on the other, by the operating system.

In my research, I studied how all these different actors trust each other, and how this trust can be unfortunately violated by attackers, due to limitations on how the mobile operating systems, apps, and user interfaces are currently designed and implemented. To assist my work, I developed automated systems to perform large-scale analyses of mobile apps.

In this talk, I will describe both the tools I have developed and my findings. Specifically, I will first describe my work on how, in an Android system, it is possible to lure users to interact with malicious apps which “look like” legitimate ones. This attack completely violates the trust relationship, mediated by the user interface, between users and apps. Then, I will explain how many apps unsafely authenticate their users to remote backends, due to misplaced trust in the operating system. Finally, I will show how many apps misuse hardware-backed authentication devices, such as trusted execution environments and fingerprint readers, making them vulnerable to a variety of authentication bypass attacks. I will finish my talk presenting current open issues in the field and outlining future directions for my research.

Speaker Biography: Antonio Bianchi is a Ph.D. candidate at University of California, Santa Barbara (UCSB). His main research interest is in the area of Computer Security, with a focus on Mobile Systems. During his Ph.D., he worked on discovering, studying, and fixing novel security issues in the ecosystem of mobile devices and applications. He also explored research interests in other fields of computer security, such as binary program analysis and hardening, hardware-assisted authentication, and security of the Internet of Things.

Video Recording >>

March 1, 2018

Datacenters host a wide range of today’s low-latency applications. To meet their strict latency requirements at scale, datacenter networks are designed as topologies that can provide a large number of parallel paths between each pair of hosts. The recent trend towards simple datacenter network fabric strips most network functionality, including load balancing among these paths, out of the network core and pushes it to the edge. This slows reaction to microbursts, the main culprit of packet loss — and consequently performance degradation — in datacenters. We investigate the opposite direction: could slightly smarter fabric significantly improve load balancing? I will present DRILL, a datacenter fabric which performs micro load balancing to distribute load as evenly as possible on microsecond timescales. DRILL employs per-packet decisions at each switch based on local queue occupancies and randomized algorithms to distribute load. I will explain how we address the resulting key challenges of packet reordering and topological asymmetry and present results showing that DRILL outperforms recent edge-based load balancers, particularly under heavy load while imposing only minimal (less than 1%) switch area overhead. Under 80% load, for example, it achieves 1.3-1.4x lower mean flow completion time than recent proposals. Finally, I will discuss our analysis of DRILL’s stability and throughput-efficiency. I will conclude by discussing some of the challenges and opportunities of applying cross-layer micro load balancing to improve the performance of some critical network applications.

Speaker Biography: Soudeh Ghorbani is a researcher in computer networks. She received her PhD from the University of Illinois at Urbana-Champaign in 2016 advised by Brighten Godfrey, and during 2017 was a postdoctoral research associate at the University of Wisconsin working with Aditya Akella. Her research has won a number of awards and fellowships including the VMware Graduate Fellowship (one of the 3 winners in 2015, worldwide), the best paper award at HotSDN, the Feng Chen Memorial Award, and the Gottlieb Fellowship.


March 5, 2018

Robots are increasingly an important part of our world and our economy, and are being trusted with more and more complex tasks while working alongside and interacting with ordinary people. As such, it is particularly important that robots can be taught new skills on the fly by those who are not robotics experts, but are instead experts in their domain. Enabling this requires a wide variety of tools including new user interfaces, methods for learning tasks from demonstration, and new algorithms for intelligent execution of skills.

In this work, we discuss how we can allow end users to create more complex task plans incorporating reasoning and perception through the Behavior Tree-based CoSTAR interface. CoSTAR is the 2016 KUKA Innovation Award-winning interface for collaborative robot instruction. A 35-person study with novice users showed that CoSTAR was an effective and usable system for authoring robot task plans.

However, these plans are limited in how they can adapt to new environments on their own. New algorithms that combine perceptual abstractions with learning and motion planning allow the robot to better adaptation to new environments and to new tasks. In particular, we see that the combination of learned skill models with tree search allows for robust adaptation to new environments. The secret to humanlike generalization is to combine low-level motion policies with high-level task planning; it amounts to giving the robots some measure of common sense.

Together, these components allow for powerful behaviors that can adapt to new environments. Finally, we explore how the robot can generalize and reason about these high-level task decisions by using learned models. These models allow us to combine and execute lower-level behaviors by giving the robot an “imagination” which means it can predict the effects of any individual policy on the world.

This combination of user interfaces, imitation learning, and new planning and machine learning algorithms will allow ordinary end users to create more complex, powerful, and understandable task plans for collaborative robots.

Speaker Biography: Chris Paxton received his B.S. in Computer Science with a minor in Neuroscience from the University of Maryland, College Park, in 2012. He then came to the Johns Hopkins University in Baltimore, MD, where he was given the Intuitive Surgical fellowship from 2012-2014. His first project was using machine learning with electronic medical records for early prediction of septic shock before switching focus to creating and using task plans for collaborative robots. He worked on the original version of CoSTAR and led the team that won the 2016 KUKA Innovation Award in Hannover, Germany. Chris is interested in how we can let robots solve problems the way people do, both because it will help us build more useful systems and because it tells us something about ourselves. He plans to continue his research post-graduation as a research scientist with NVidia.

Video Recording >>

March 6, 2018

Machine learning (ML) systems are increasingly deployed in safety- and security-critical domains such as self-driving cars and malware detection, where the system correctness for corner case inputs are crucial. Existing testing of ML system correctness depends heavily on manually labeled data and therefore often fails to expose erroneous behaviors for rare inputs.

In this talk, I will present the first framework to test and repair ML systems, especially in an adversarial environment. In the first part, I will introduce DeepXplore, a whitebox testing framework of real-world deep learning (DL) systems. Our evaluation shows that DeepXplore can successfully find thousands of erroneous corner case behaviors, e.g., self-driving cars crashing into guard rails and malware masquerading as benign software. In the second part, I will introduce machine unlearning, a general, efficient approach to repair an ML system exhibiting erroneous behaviors. Our evaluation, on four diverse learning systems and real-world workloads, shows that machine unlearning is general, effective, fast, and easy to use.

Speaker Biography: Yinzhi Cao is an assistant professor at Lehigh University. He earned his Ph.D. in Computer Science at Northwestern University and worked at Columbia University as a postdoc. Before that, he obtained his B.E. degree in Electronics Engineering at Tsinghua University in China. His research mainly focuses on the security and privacy of the Web, smartphones, and machine learning. He has published many papers at various security and system conferences, such as IEEE S&P (Oakland), NDSS, CCS, and SOSP. His JShield system has been adopted by Huawei, the world’s largest telecommunication company. His past work was widely featured by over 30 media outlets, such as NSF Science Now (Episode 38), CCTV News, IEEE Spectrum, Yahoo! News and ScienceDaily. He received two best paper awards at SOSP’17 and IEEE CNS’15 respectively. He is one of the recipients of 2017 Amazon Research Awards (ARA).

Video Recording >>

March 8, 2018

In the past decade there has been a significant increase in the collection of personal information and communication metadata (with whom users communicate, when, how often) by governments, Internet providers, companies, and universities. While there are many ongoing efforts to secure users’ communications, namely end-to-end encryption messaging apps and E-mail services, safeguarding metadata remains elusive.

I will present a system called Pung that makes progress on this front. Pung lets users exchange messages over the Internet without revealing any information in the process. Perhaps surprisingly, Pung achieves this strong privacy property even when all providers (ISPs, companies, etc.) are arbitrarily malicious.

I will also present several improvements to a general cryptographic building block called private information retrieval (PIR) that underlies many privacy preserving systems including Pung. Among these improvements, I will discuss SealPIR, a new PIR library that achieves orders of magnitude more network efficiency than the state-of-the-art. Finally, I will briefly touch on some of my work on verifiable computation and interfacing with malicious USB devices.

Speaker Biography: Sebastian Angel is a Ph.D. candidate at The University of Texas at Austin and a visiting academic at New York University’s Courant Institute of Mathematical Sciences. His research interests are in systems, security, and networking.


March 12, 2018

Hardware trends over the last several decades have lead to shifting priorities with respect to performance bottlenecks in the implementations of dataflows typically present in large-scale data analytics applications. In particular, efficient use of main memory has emerged as a critical aspect of dataflow implementation, due to the proliferation of multi-core architectures, as well as the rapid development of faster-than-disk storage media. At the same time, the wealth of static domain-specific information about applications remains an untapped resource when it comes to optimizing the use of memory in a dataflow application.

We propose a compilation-based approach to the synthesis of memory-efficient dataflow implementations, using static analysis to extract and leverage domain-specific information about the application. Our program transformations use the combined results of type, effect, and provenance analyses to infer time- and space- effective placement of primitive memory operations, precluding the need for dynamic memory management and its attendant costs. The experimental evaluation of implementations synthesized with our framework shows both the importance of optimizing for memory performance, as well as significant benefits of our approach, along multiple dimensions.

Finally, we also demonstrate a framework for formally verifying the soundness of these transformations, laying the foundation for their use as a component of a more general implementation synthesis ecosystem.

Speaker Biography: Shyam is a PhD candidate in Computer Science at JHU, and a member of the Data Management and Systems Lab (DaMSL) under the supervision of Dr. Yanif Ahmad. Over the course of his PhD, Shyam’s work and interests have spanned the gamut from the most arcane of theory, to the depths of systems performance engineering.

After wrapping up at Hopkins, Shyam will be joining Galois — a private research consultancy — as a member of the technical staff.

March 14, 2018

Artificial intelligence has begun to impact healthcare in areas including electronic health records, medical images, and genomics. But one aspect of healthcare that has been largely left behind thus far is the physical environments in which healthcare delivery takes place: hospitals and assisted living facilities, among others. In this talk I will discuss my work on endowing hospitals with ambient intelligence, using computer vision-based human activity understanding in the hospital environment to assist clinicians with complex care. I will first present an implementation of an AI-Assisted Hospital where we have equipped units at two partner hospitals with visual sensors. I will then discuss my work on human activity understanding, a core problem in computer vision. I will present deep learning methods for dense and detailed recognition of activities, and efficient action detection, important requirements for ambient intelligence. I will discuss these in the context of two clinical applications, hand hygiene compliance and automated documentation of intensive care unit activities. Finally, I will present work and future directions for integrating this new source of healthcare data into the broader clinical data ecosystem, towards full realization of an AI-Assisted Hospital.

Speaker Biography: Serena Yeung is a PhD candidate at Stanford University in the Artificial Intelligence Lab, advised by Fei-Fei Li and Arnold Milstein. Her research focuses on deep learning and computer vision algorithms for video understanding and human activity recognition. More broadly, she is passionate about using these algorithms to equip healthcare spaces with ambient intelligence, in particular an AI-Assisted Hospital. Serena is the lead graduate student in the Stanford Partnership in AI-Assisted Care (PAC), a collaboration between the Stanford School of Engineering and School of Medicine. She interned at Facebook AI Research in 2016, and Google Cloud AI in 2017. She was also co-instructor for Stanford’s CS231N course on Convolutional Neural Networks for Visual Recognition in 2017.

Video Recording >>

March 15, 2018

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.

Speaker Biography: 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.


March 22, 2018

Evaluating anatomical variations in structures like the nasal passage and sinuses is challenging because their complexity can often make it difficult to differentiate normal and abnormal anatomy. By statistically modeling these variations and estimating individual patient anatomy using these models, quantitative estimates of similarity or dissimilarity between the patient and the sample population can be made. In order to do this, a spatial alignment, or registration, between patient anatomy and the model must first be computed. In this dissertation, a deformable most likely point paradigm is introduced that incorporates statistical variations into feature-based registration algorithms. This paradigm is a variant of the most likely point paradigm, which incorporates feature uncertainty into the registration process. Our deformable registration algorithms optimize the probability of feature alignment as well as the probability of model deformation allowing statistical models of anatomy to estimate, for instance, structures seen in endoscopic video without the need for patient specific computed tomography (CT) scans. The probabilistic framework also enables the algorithms to assess the quality of registrations produced, allowing users to know when an alignment can be trusted. This talk will cover 3 algorithms built within this paradigm and evaluated in simulation and in-vivo experiments.

Speaker Biography: Ayushi is a PhD candidate in Computer Science at JHU under the supervision of Dr. Russ Taylor and Dr. Greg Hager. She received her B.S in Computer Science and B.A. in Mathematics from Providence College, RI in 2011. During the course of her PhD, she worked on improving statistical shape models of anatomy and on using these models in deformable registration techniques. After finishing her PhD, Ayushi plans to continue developing these ideas further as a Provost Postdoctoral Fellow at JHU.

Video Recording >>

March 22, 2018

Companies such as Google or Lyft collect a substantial amount of location data about their users to provide useful services. The release of these datasets for general use can enable numerous innovative applications and research. However, such data contains sensitive information about the users, and simple clocking-based techniques have been shown to be ineffective to ensure users’ privacy. These privacy concerns have motivated many leading technology companies and researchers to develop algorithms that collect and analyze location data with formal provable privacy guarantees. I will show a unified framework that can (a) enhance a better understanding about the many existing provable privacy guarantees for location data; (b) allow flexible trade-offs between privacy, accuracy, and performance, based on the application’s requirements. I will also describe some exciting new research about provable privacy guarantees for handling advanced settings involving complex queries or datasets and emerging data-driven applications, and conclude with directions for future privacy research in big-data management and analysis.

Speaker Biography: Xi He is a Ph.D. student at Computer Science Department, Duke University. Her research interests lie in privacy-preserving data analysis and security. She has also received a double degree in Applied Mathematics and Computer Science from the University of Singapore. Xi has been working with Prof. Machanavajjhala on privacy since 2012. She has published in SIGMOD, VLDB, and CCS, and has given tutorials on privacy at VLDB 2016 and SIGMOD 2017. She received best demo award on differential privacy at VLDB 2016 and was awarded a 2017 Google Ph.D. Fellowship in Privacy and Security.

Video Recording >>

March 29, 2018

Over the next few decades, we are going to transition to a new economy where highly complex, customizable products are manufactured on demand by flexible robotic systems. In many fields, this shift has already begun. 3D printers are revolutionizing production of metal parts in the aerospace, automotive, and medical industries. Whole-garment knitting machines allow automated production of complex apparel and shoes. Manufacturing electronics on flexible substrates makes it possible to build a whole new range of products for consumer electronics and medical diagnostics. Collaborative robots, such as Baxter from Rethink Robotics, allow flexible and automated assembly of complex objects. Overall, these new machines enable batch-one manufacturing of products that have unprecedented complexity.

In my talk, I argue that the field of computational design is essential for the next revolution in manufacturing. To build increasingly functional, complex and integrated products, we need to create design tools that allow their users to efficiently explore high-dimensional design spaces by optimizing over a set of performance objectives that can be measured only by expensive computations. I will discuss how to overcome these challenges by 1) developing data-driven methods for efficient exploration of these large spaces and 2) performance-driven algorithms for automated design optimization based on high-level functional specifications. I will showcase how these two concepts are applied by developing new systems for designing robots, drones, and furniture. I will conclude my talk by discussing open problems and challenges for this emerging research field.

Speaker Biography: Adriana Schulz is a Ph.D. student in the Department of Electrical Engineering and Computer Science at MIT where she works at the Computer Science and Artificial Intelligence Laboratory. She is advised by Professor Wojciech Matusik and her research spans computational design, digital manufacturing, interactive methods, and robotics. Before coming to MIT, she obtained a M.S. in mathematics from IMPA, Brazil and a B.S. in electronics engineering from UFRJ, Brazil.


March 29, 2018

Parsing object semantics and geometry in a scene is one core task in visual understanding. This includes localization of an object, classification of its identity, estimation of object orientation and parsing 3D shape structures. With the emergence of deep convolutional architectures in recent years, substantial progress has been made for large-scale vision problems such as image classification. However, there still remains some fundamental challenges. First, creating object representations that are robust to changes in viewpoint while capturing local visual details continues to be a problem. Second, deep Convolutional Neural Networks (CNNs) are purely driven by data and predominantly pose the scene interpretation problem as an end-to-end black-box mapping. However, decades of work on perceptual organization in both human and machine vision suggests that there are often intermediate representations that are intrinsic to an inference task, and which provide essential structure to improve generalization.

In this dissertation, we present two methodologies to surmount the aforementioned two issues. We first introduce a multi-domain pooling framework which group local visual signals within generic feature spaces that are invariant to 3D object transformation, thereby reducing the sensitivity of output feature to spatial deformations. Next, we explore an approach for injecting prior domain structure into neural network training, which leads a CNN to recover a sequence of intermediate milestones towards the final goal. We implement this deep supervision framework with a novel CNN architecture which is trained on synthetic image only and achieves the state-of-the-art performance of 2D/3D keypoint localization on real image benchmarks. Finally, the proposed deep supervision scheme also motivates an approach for accurately inferring six Degree-of-Freedom (6-DoF) pose for a large number of object classes from single or multiple views.

Speaker Biography: Chi Li is a Ph.D. candidate primarily advised by Dr. Greg Hager. He received his B.E. from Cognitive Science Department at Xiamen University in 2012, where he became interested in computer vision. His research mainly focuses on visual understanding of object properties from semantic class to 3D pose and structure. In particular, he is interested in leveraging scene geometry to enhance deep learning techniques on 2D/3D/Multi-view perception. During his Ph.D., he also gain industrial experience from his three research internships with Apple, NEC Laboratories America and Microsoft Research.

Professional plans after Hopkins: I am going to join Apple and continue my research of 2D/3D visual perception after graduation.

Video Recording >>

April 10, 2018

Just like programming a robot requires meticulous planning, coding, and execution, these same requirements are ever present when designing and controlling the individual optical and acoustic components of photoacoustic imaging systems. Photoacoustic imaging utilizes light and sound to make images by transmitting laser pulses that illuminate regions of interest, which subsequently absorb the light, causing thermal expansion and the generation of sound waves that are detected with conventional ultrasound transducers. The Photoacoustic and Ultrasonic Systems Engineering (PULSE) Lab is developing novel methods that use photoacoustic imaging to guide surgeries with the ultimate goal of eliminating surgical complications caused by injury to important structures – like major blood vessels and nerves – that are otherwise hidden from a surgeon’s immediate view.

In this talk, I will describe our novel light delivery systems that attach to surgical tools in order to direct light toward the surgical site. I will also introduce how we learn from the physics of sound propagation in tissue to develop acoustic beamforming algorithms that improve image quality, using both state-of-the-art deep learning methods and our newly developed spatial coherence theory. These light delivery and acoustic beamforming methods hold promise for robotic tracking tasks, visualization and visual servoing of surgical tool tips, and assessment of relative distances between the surgical tool and nearby critical structures (e.g., major blood vessels and nerves that if injured will cause severe complications, paralysis, or patient death). Impacted surgeries and procedures include neurosurgery, spinal fusion surgery, hysterectomies, and biopsies.

Speaker Biography: Muyinatu Bell is an Assistant Professor of Electrical and Computer Engineering with a joint appointment in the Biomedical Engineering Department at Johns Hopkins University, where she founded and directs the Photoacoustic and Ultrasonic Systems Engineering (PULSE) Lab. Dr. Bell earned a B.S. degree in Mechanical Engineering (biomedical engineering minor) from Massachusetts Institute of Technology (2006), received a Ph.D. degree in Biomedical Engineering from Duke University (2012), and conducted research abroad as a Whitaker International Fellow at the Institute of Cancer Research and Royal Marsden Hospital in the United Kingdom (2009-2010). Prior to joining the faculty, Dr. Bell completed a postdoctoral fellowship with the Engineering Research Center for Computer-Integrated Surgical Systems and Technology at Johns Hopkins University (2016), where she was co-mentored by faculty in the Computer Science Department and the School of Medicine. Dr. Bell has published over 40 scientific journal articles and conference papers, holds a patent for short-lag spatial coherence beamforming, and is the recipient of numerous awards, grants, and fellowships, including the NIH K99/R00 Pathway to Independence Award (2015), MIT Technology Review’s Innovator Under 35 Award (2016), and the NSF CAREER Award (2018).


April 11, 2018

Single population, biologically-inspired algorithms such as Genetic Algorithm and Particle Swarm Optimization are effective tools for solving a variety of optimization problems. Like many such algorithms, however, they fall victim to the curse of dimensionality. Additionally, these algorithms often suffer from a phenomenon known as hitchhiking where improved solutions are not unequivocally better for all variables. Insofar as individuals within these populations are deemed to be competitive, one solution to both the curse of dimensionality and the problem of hitchhiking has been to introduce more cooperation. These multi-population algorithms cooperate by decomposing a problem into parts and assigning a population to each part.

Factored Evolutionary Algorithms (FEA) generalize this decomposition and cooperation to any evolutionary algorithm. A key element of FEA is a global solution that provides missing information to individual populations and coordinates them. This dissertation extends FEA to the distributed case by having individual populations maintain and coordinate local solutions that maintain consensus. This Distributed FEA (DFEA) is demonstrated to perform well on a variety of problems and, sometimes, even if consensus is lost. However, DFEA fails to maintain the same semantics as FEA.

To address this issue, we develop an alternative framework to the “cooperation versus competition” dichotomy. In this framework, information flows are modeled as a blackboard architecture. Changes in the blackboard are modeled as merge operations that require conflict resolution between existing and candidate values. Conflict resolution is handled using Pareto efficiency, which avoids hitchhiking. We apply this framework to FEA and DFEA and develop revised DFEA, which performs identically to FEA.

We then apply our framework to a single population algorithm, Particle Swarm Optimization (PSO), to create Pareto Improving PSO (PI-PSO). We demonstrate that PI-PSO outperforms PSO and sometimes FEA-PSO, often with fewer individuals.

Finally, we extend our information based approach by implementing parallel, distributed versions of FEA and DFEA using the Actor model. The Actor model is based on message passing, which accords well with our information-centric framework. We use validation experiments to verify that we have successfully implemented the semantics of the serial versions of FEA and DFEA

Speaker Biography: Butcher is a PhD candidate in Computer Science at JHU, and a member of Numerical Intelligent Systems Laboratory (NISL) under the direction of Dr. John Sheppard. He has broad interests in software engineering, artificial intelligence and statistics. He works as a software engineer/data scientist for Appriss Safety. He also teaches Artificial Intelligence and Data Science at JHU Whiting School of Engineering’s Engineering for Professionals program. He is a Zen priest.

After wrapping up at Hopkins, Butcher will continue to working and teaching…and doing a bit of Zen.

Distinguished Lecturer

April 12, 2018

We research on autonomous mobile robots with a seamless integration of perception, cognition, and action. In this talk, I will first introduce our CoBot service robots and their novel localization and symbiotic autonomy, which enable them to consistently move in our buildings, and learn from asking humans or the web for help to overcome their limitations. I frame the research as human-AI interaction also including an interpretability approach by language generation to alert and respond to human explanation requests. I will conclude with a brief discussion of AI, machine learning, and robotics, and their potential social impact.

Speaker Biography: Manuela M. Veloso is Herbert A. Simon University Professor in the School of Computer Science at Carnegie Mellon University. Currently, she is the Head of the Machine Learning Department. She researches in Artificial Intelligence, Robotics, and Machine Learning. She founded and directs the CORAL research laboratory, for the study of autonomous agents that Collaborate, Observe, Reason, Act, and Learn, www.cs.cmu.edu/~coral. Professor Veloso is ACM Fellow, IEEE Fellow, AAAS Fellow, AAAI Fellow, Einstein Chair Professor, the co-founder and past President of RoboCup, and past President of AAAI. Professor Veloso and her students research with a variety of autonomous robots, including mobile service robots and soccer robots. See www.cs.cmu.edu/~mmv for further information, including publications.

Video Recording >>

April 17, 2018

Web applications are integral to today’s society, hosting a variety of services ranging from banking and e-commerce to mapping and social media. To support these rich services, web applications have evolved into complex distributed systems, making critical tasks such as performance optimization and debugging difficult.

In this talk, I will describe how we can address this growing complexity by efficiently identifying and analyzing the fine-grained, distributed data flows in web applications. Tracking data flows at the granularity of individual pieces of program state, like JavaScript variables on the client-side, and key/value pairs in storage systems on the server-side, provides invaluable insights into the low-level behavior of complex web services. This information enables a variety of systems with new, more powerful performance optimizations and debugging techniques. I will describe two such systems that we have built. The first is Polaris, a web page load optimizer that identifies data dependencies between web objects to improve browser request scheduling and reduce page load times by 34%-59%. I will then discuss Vesper, the first system to accurately measure how quickly web pages become interactive for users. Vesper uses fine-grained data flows to automatically identify a page’s interactive state and reduce page time-to-interactivity by 32%. I will conclude by discussing some future research challenges involving large-scale web services.

Speaker Biography: Ravi Netravali is a Ph.D. student at MIT, advised by Professors Hari Balakrishnan and James Mickens. His research interests are in computer systems and networks, with a recent focus on building practical systems to improve the performance and debugging of large-scale, distributed web applications. He is a recipient of the 2017 Qualcomm Innovation Fellowship, and shared the Internet Research Task Force’s Applied Networking Research Prize in 2018. Netravali graduated from Columbia University in 2012 with a B.S. in Electrical Engineering.

Video Recording >>

April 19, 2018

The ongoing explosion of spatiotemporal tracking data has now made it possible to analyze and model fine-grained behaviors in a wide range of domains. For instance, tracking data is now being collected for every NBA basketball game with players, referees, and the ball tracked at 25 Hz, along with annotated game events such as passes, shots, and fouls. Other settings include laboratory animals, people in public spaces, professionals in settings such as operating rooms, actors speaking and performing, digital avatars in virtual environments, and even the behavior of other computational systems.

Motivated by these applications, I will describe recent and ongoing work in developing principled structured imitation learning approaches that can exploit interdependencies in the state/action space, and achieve orders-of-magnitude improvements compared to conventional approaches in learning rate or accuracy, or both. These approaches are showcased on a wide range of (often commercially deployed) applications, including modeling professional sports, laboratory animals, speech animation, and expensive computational oracles.

Speaker Biography: Yisong Yue is an assistant professor in the Computing and Mathematical Sciences Department at the California Institute of Technology. He was previously a research scientist at Disney Research. Before that, he was a postdoctoral researcher in the Machine Learning Department and the iLab at Carnegie Mellon University. He received a Ph.D. from Cornell University and a B.S. from the University of Illinois at Urbana-Champaign.

Yisong’s research interests lie primarily in the theory and application of statistical machine learning. His research is largely centered around developing integrated learning-based approaches that can characterize complex structured and adaptive decision-making settings. Current focus areas include developing novel methods for spatiotemporal reasoning, structured prediction, interactive learning systems, and learning with humans in the loop. In the past, his research has been applied to information retrieval, recommender systems, text classification, learning from rich user interfaces, analyzing implicit human feedback, data-driven animation, behavior analysis, sports analytics, policy learning in robotics, and adaptive routing & allocation problems.

Video Recording >>

Distinguished Lecturer

April 26, 2018

For the past six years, the Google Brain team (g.co/brain) has conducted research on difficult problems in artificial intelligence, on building large-scale computer systems for machine learning research, and, in collaboration with many teams at Google, on applying our research and systems to dozens of Google products. Our group has open-sourced the TensorFlow system (tensorflow.org), a widely popular system designed to easily express machine learning ideas, and to quickly train, evaluate and deploy machine learning systems. In this talk, I’ll highlight some of the design decisions we made in building TensorFlow, discuss research results produced within our group in areas such as computer vision, language understanding, translation, healthcare, and robotics, and describe ways in which these ideas have been applied to a variety of problems in Google’s products, usually in close collaboration with other teams. I will also touch on some exciting areas of research that we are currently pursuing within our group.

This talk describes joint work with many people at Google.

Speaker Biography: Jeff Dean (research.google.com/people/jeff) joined Google in 1999 and is currently a Google Senior Fellow in Google’s Research Group, where he co-founded and leads the Google Brain team, Google’s deep learning and artificial intelligence research team. He and his collaborators are working on systems for speech recognition, computer vision, language understanding, and various other machine learning tasks. He has co-designed/implemented many generations of Google’s crawling, indexing, and query serving systems, and co-designed/implemented major pieces of Google’s initial advertising and AdSense for Content systems. He is also a co-designer and co-implementor of Google’s distributed computing infrastructure, including the MapReduce, BigTable and Spanner systems, protocol buffers, the open-source TensorFlow system for machine learning, and a variety of internal and external libraries and developer tools.

Jeff received a Ph.D. in Computer Science from the University of Washington in 1996, working with Craig Chambers on whole-program optimization techniques for object-oriented languages. He received a B.S. in computer science & economics from the University of Minnesota in 1990. He is a member of the National Academy of Engineering, and of the American Academy of Arts and Sciences, a Fellow of the Association for Computing Machinery (ACM), a Fellow of the American Association for the Advancement of Sciences (AAAS), and a winner of the ACM Prize in Computing.


April 27, 2018

Over the last decade, American hospitals have adopted electronic health records (EHRs) widely. In the next decade, incorporating EHRs with clinical decision support (CDS) together into the process of medicine has the potential to change the way medicine has been practiced and advance the quality of patient care. It is a unique opportunity for machine learning (ML), with its ability to process massive datasets beyond the scope of human capability, to provide new clinical insights that aid physicians in planning and delivering care, ultimately leading to better outcomes, lower costs of care, and increased patient satisfaction. However, applying ML-based CDS has to face steep system and application challenges. No open platform is there to support ML and domain experts to develop, deploy, and monitor ML-based CDS; and no end-to-end solution is available for machine learning algorithms to consume heterogenous EHRs and deliver CDS in real-time. Build ML-based CDS from scratch can be expensive and time-consuming.

In this dissertation, CDS-Stack, an open cloud-based platform, is introduced to help ML practitioners to deploy ML-based CDS into healthcare practice. The CDS-Stack integrates various components into the infrastructure for the development, deployment, and monitoring of the ML-based CDS. It provides an ETL engine to transform heterogenous EHRs, either historical or online, into a common data model (CDM) in parallel so that ML algorithms can directly consume health data for training or prediction. It introduces both pull and push-based online CDS pipelines to deliver CDS in real-time. The CDS-Stack has been adopted by Johns Hopkins Medical Institute (JHMI) to deliver a sepsis early warning score since November 2017 and begins to show promising results. Furthermore, we believe CDS-Stack can be extended to outpatients too. A case study of outpatient CDS has been conducted which utilizes smartphones and machine learning to quantify the severity of Parkinson disease. In this study, a mobile Parkinson disease severity score (mPDS) is generated using a novel machine learning approach. The results show it can detect response to dopaminergic therapy, correlate strongly with traditional rating scales, and capture intraday symptom fluctuation.

Speaker Biography: Andong Zhan is a Ph.D. candidate in Computer Science whose mission is to build novel computer systems with machine learning to solve challenging healthcare problems. He has designed and deployed several computing systems for healthcare which directly influenced thousands of inpatients and outpatients as well. Thus far, he has published more than 20 peer-reviewed articles in the areas of computer science and clinical research including IEEE Communications, ACM SenSys, JAMA Neurology, Parkinsonism & Related Disorders, and Critical Care Medicine. His Parkinson disease remote monitoring project and mobile Parkinson disease severity score have been reported by Nature Outlook 2016, JHU Hub, medpagetoday.com, etc.; meanwhile, a machine learning based clinical decision system he mainly developed keeps monitoring inpatients in Johns Hopkins Hospitals and generating real-time alerts for those who may develop sepsis. He is going to graduate in May 2018 and is looking for his next adventure. His advisers are Prof. Andreas Terzis and Prof. Suchi Saria. Also, Prof. Yanif Ahmad gave him massive guidance on his research. He received his M.E from the Dept. of Computer Science at Nanjing University, Nanjing, China, in June 2010 and his B.S from the Dept. of Computer Science at Nanjing University, Nanjing, China, in June 2007.


April 27, 2018

Robustness has always been a desirable property for natural language processing (NLP). In many cases, NLP models (e.g., parsing) and downstream applications (e.g., machine translation) perform poorly when the input contains noise such as spelling errors, grammatical errors, and disfluency. In this thesis, I present robust error correction models for language learners’ texts at different levels of granularity: character, token, and sentence-level errors.

Character and token-level errors in language learners’ writing are related to grammar, and NLP community has focused on these error types for a long time. I review prior work (particularly focusing in the last ten years) on grammatical error correction (GEC), and present new models for character and token-level error correction. For character level, I introduce a semi-character recurrent neural network, which is motivated by a finding in Psycholinguistics, called Cmabrigde Uinervtisy (Cambridge University) effect. For word-level robustness, I propose an error-repair dependency parsing algorithm for ungrammatical texts. The algorithm can parse sentences and correct grammatical errors simultaneously.

NLP community has also extended the scope of errors to phase and sentence-level errors, where fluency comes into play as an important notion. This extension of the scope and the notion of fluency bring us new challenges for GEC such as evaluation metrics and data annotation. After I introduce a method for efficient human judgment collection using bayesian online updates, I present a new annotation scheme and dataset for sentence-level GEC, followed by a neural encoder-decoder GEC model that directly optimizes toward a metric to avoid exposure bias.

Finally, I conclude the thesis and outline ideas and suggestions for future research in GEC.

Speaker Biography: Keisuke Sakaguchi is a Ph.D. candidate advised by Benjamin Van Durme and Matt Post in the department of Computer Science, Center of Language and Speech Processing at Johns Hopkins University. His research focuses upon robust natural language processing (NLP) for ungrammatical noisy texts and NLP for educational purposes. He has received an Outstanding Paper Award at ACL 2017. He received his M.Eng. in Information Science at Nara Institute of Science and Technology, M.A. in Psycho&Neurolinguistics at University of Essex, and B.A. in Literature (major in Philosophy) at Waseda University.

Video Recording >>

May 1, 2018

Elliot Soloway’s Rainfall problem is a classic benchmark in computing education research, designed to study plan composition: how students decompose problems into tasks, solve them, and compose the solutions.

Over multiple decades, locations, and languages, students have done poorly at it. However, recent multi-institution studies of students using functional programming and How to Design Programs find very different outcomes.

What do these results tell us? This talk explores relationships between programming languages, program design, curricula, and how students perceive code structure.

The talk assumes no experience with plan composition, functional programming, or having been rained upon. Please come equipped with pen and paper, because the talk will require you to write programs.

Joint work primarily with Kathi Fisler.

Speaker Biography: Shriram Krishnamurthi is a Professor of Computer Science and an Associate Director of the Executive Master in Cybersecurity at Brown University. With collaborators and students, he has created several influential systems: DrRacket and WeScheme (programming environments), Margrave (policy analyzer), FrTime and Flapjax (reactive programming languages), Lambda-JS and TeJaS (semantics and types for JavaScript), and Flowlog (software-defined networking programming language and verifier). He is now working on the Pyret programming language. He is the author of “Programming Languages: Application and Interpretation” and a co-author of “How to Design Programs” and “Programming and Programming Languages”. He co-directs the Bootstrap math-and-computing outreach program. He won SIGPLAN’s Robin Milner Young Researcher Award, SIGSOFT’s Influential Educator Award, and Brown’s Henry Merritt Wriston Fellowship for distinguished contribution to undergraduate education. He has authored over a dozen papers recognized for honors by program committees. He has an honorary doctorate from the Università della Svizzera Italiana.

Video Recording >>

May 3, 2018

Medical devices, autonomous vehicles, and the Internet of Things depend on the integrity and availability of trustworthy data from sensors to make safety-critical, automated decisions. How can such cyberphysical systems remain secure against an adversary using intentional interference to fool sensors? Building upon classic research in cryptographic fault injection and side channels, research in analog cybersecurity explores how to protect digital computer systems from physics-based attacks. Analog cybersecurity risks can bubble up into operating systems as bizarre, undefined behavior. For instance, transduction attacks exploit vulnerabilities in the physics of a sensor to manipulate its output. Transduction attacks using audible acoustic, ultrasonic, or radio interference can inject chosen signals into sensors found in devices ranging from fitbits to implantable medical devices to drones and smartphones.

Why do microprocessors blindly trust input from sensors, and what can be done to establish trust in unusual input channels in cyberphysical systems? Why are students taught to hold the digital abstraction as sacrosanct and unquestionable? Come to this talk to learn about undefined behavior in basic building blocks of computing. I will discuss how to design out analog cybersecurity risks by rethinking the computing stack from electrons to bits. I will also suggest educational opportunities for embedded security and the role of tenure for interdisciplinary federal engagement beyond an individual’s own research agenda.

Speaker Biography: Kevin Fu is Associate Professor of EECS at the University of Michigan where he directs the SPQR lab (SPQR.eecs.umich.edu) and the Archimedes Center for Medical Device Security (secure-medicine.org). His research focuses on analog cybersecurity—how to model and defend against threats to the physics of computation and sensing. His embedded security research interests span from the physics of cybersecurity through the operating system to human factors. Past research projects include MEMS sensor security, pacemaker/defibrillator security, cryptographic file systems, web authentication, RFID security and privacy, wirelessly powered sensors, medical device safety, and public policy for information security & privacy.

Kevin was recognized as an IEEE Fellow, Sloan Research Fellow, MIT Technology Review TR35 Innovator of the Year, and recipient of a Fed100 Award and NSF CAREER Award. He received best paper awards from USENIX Security, IEEE S&P, and ACM SIGCOMM. He co-founded healthcare cybersecurity startup Virta Labs. Kevin has testified in the House and Senate on matters of information security and has written commissioned work on trustworthy medical device software for the National Academy of Medicine. He is a member the Computing Community Consortium Council, ACM Committee on Computers and Public Policy, and the USENIX Security Steering Committee. He advises the American Hospital Association and Heart Rhythm Society on matters of healthcare cybersecurity. Kevin previously served as program chair of USENIX Security, a member of the NIST Information Security and Privacy Advisory Board, a visiting scientist at the Food & Drug Administration, and an advisor for Samsung’s Strategy and Innovation Center. Kevin received his B.S., M.Eng., and Ph.D. from MIT. He earned a certificate of artisanal bread making from the French Culinary Institute.

Computer Science Student Defense

June 12, 2018

This talk and its corresponding dissertation address the problem of learning video representations, which is defined here as transforming the video so that its essential structure is made more visible or accessible for action recognition and quantification. In the literature, a video can be represented by a set of images, by modeling motion or temporal dynamics, and by a 3D graph with pixels as nodes. This dissertation contributes in proposing a set of models to localize, track, segment, recognize and assess actions such as (1) image set via aggregating subset features given by regularizing normalized Convolutional Neural Networks (CNNs), (2) image set via inter-frame principal recovery and sparsely coding residual actions, (3) temporally local models with spatially global motion estimated by robust feature matching and local motion estimated by action detection with motion model added, (4) spatiotemporal models with actions segmented by 3D Graph cuts and quantified by segmental 3D CNNs, respectively.

The state-of-the-art performances have been achieved for tasks such as quantifying actions of the facial pain and human diving. The primary conclusions of this dissertation are categorized as follows: (i) Image set models can effectively capture facial actions that are about collective representation; (ii) The sparse and low-rank representations of facial actions can have the expression, identity and poses cues untangled and can be learned using an image-set model and also a linear model; (iii) It is shown from face nets that norm is related with recognizability and that the similarity metric and loss function matter; (v) Combining the Multiple Instance Learning based boosting tracker with the Particle Filtering motion model induces a good trade-off between the appearance similarity and motion consistence; (iv) Segmenting object locally makes it amenable to assign shape priors and also it is feasible to learn knowledge such as shape priors online from the rich Web data with weak supervision; (v) It works locally in both space and time to represent videos as 3D graphs and also 3D CNNs work effectively when inputted with temporally meaningful clips.

It is hoped that the proposed models will lead to working components in building face and gesture recognition systems. In addition, the models proposed for videos can be adapted to other modalities of sequential images such as hyperspectral images and volumetric medical images which are not included in this talk and the dissertation. The model of supervised hashing by jointly learning embedding and quantization is included in the dissertation but will not be presented in the talk in the interest of time.

Speaker Biography: Xiang Xiang is a PhD student in Computer Science at Johns Hopkins University, since 2012 with Gregory D. Hager (initial appointed advisor since 2012) and a PhD candidate in Computer Science since 2014 with Trac D. Tran (primary advisor since 2014) and Gregory D. Hager (co-advisor since 2014) co-listed as my official advisors. He received the B.S. degree from the School of Computer Science at Wuhan University, Wuhan, China, in 2009, the M.S. degree from the Institute of Computing Technology at Chinese Academy of Sciences, Beijing, China, in 2012, and the M.S.E. degree from Johns Hopkins University in 2014. His research interests are computer vision and machine learning with a focus on representation learning for video understanding, facial analysis, affective computing and bio-medical applications. He has been an active member of the DSP Lab (ECE), CIRL, LCSR, MCEH, CCVL, and CIS.

Professional plans after Hopkins: Xiang Xiang will join Stefano Soatto as an Applied Research Scientist at Amazon AI.

Computer Science Student Defense:

June 18, 2018

Paraphrasing–communicating the same meaning with different surface forms–is one of the core characteristics of natural language and represents one of the greatest challenges faced by automatic language processing techniques. In this research, we investigate approaches to paraphrasing entire sentences within the constraints of a given task, which we call monolingual sentence rewriting. We focus on three representative tasks: sentence compression, text simplification, and grammatical error correction.

Monolingual rewriting can be thought of as translating between two types of English (such as from complex to simple), and therefore our approach is inspired by statistical machine translation. In machine translation, a large quantity of parallel data is necessary to model the transformations from input to output text. Parallel bilingual data naturally occurs between common language pairs (such as English and French), but for monolingual sentence rewriting, there is little existing parallel data, and annotation is costly. We modify the statistical machine translation pipeline to harness monolingual resources and insights into task constraints in order to drastically diminish the amount of annotated data necessary to train a robust system. Our method generates more meaning-preserving and grammatical sentences than earlier approaches and requires less task-specific data.

Speaker Biography: Courtney Napoles is a PhD candidate in the Computer Science Department and the Center for Language and Speech Processing at Johns Hopkins University, where she is co-advised by Chris Callison-Burch and Benjamin Van Durme. During her PhD, she interned at Educational Testing Service (ETS) and Yahoo Research. She is the recipient of an NSF Graduate Research Fellowship and holds a Bachelor’s degree in Psychology from Princeton University with a Certificate in Linguistics. Before graduate school, she edited non-fiction books for a trade publisher.

Professional plans: Courtney is a research scientist at Grammarly, where she is continuing her doctoral research.