Back to Events

Seminar Series

April 27

EECS Special Seminar: Nadia Polikarpova, "Type-Driven Program Synthesis"

Nadia Polikarpova

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-27 16:00:00 2017-04-27 17:00:00 America/New_York EECS Special Seminar: Nadia Polikarpova, "Type-Driven Program Synthesis" Abstract: Modern programming languages safeguard developers from many typical errors, yet more subtle errors—such as violations of security policies—still plague software. Program synthesis has the potential to eliminate such errors, by generating executable code from concise and intuitive high-level specifications. Traditionally, program synthesis failed to scale to specifications that encode complex behavioral properties of software: these properties are notoriously hard to check even for a given program, and so it’s not surprising that finding the right program within a large space of candidates has been considered very challenging. My work tackles this challenge through the design of synthesis-friendly program verification mechanisms, which are able to check a large set of candidate programs against a complex specification at once, whereby efficiently pruning the search space.Based on this principle, I developed Synquid, a program synthesizer that accepts specifications in the form of expressive types and uses a specialized type checker as its underlying verification mechanism. Synquid is the first synthesizer powerful enough to automatically discover provably correct implementations of complex data structure manipulations, such as insertion into Red-Black Trees and AVL Trees, and normal-form transformations on propositional formulas. Each of these programs is synthesized in under a minute. Going beyond textbook algorithms, I created a language called Lifty, which uses type-driven synthesis to automatically rewrite programs that violate information flow policies. In our case study, Lifty was able to enforce all required policies in a prototype conference management system.Bio: Nadia Polikarpova is a postdoctoral researcher at the MIT Computer Science and Artificial Intelligence Lab, interested in helping programmers build secure and reliable software. She completed her PhD at ETH Zurich. For her dissertation she developed tools and techniques for automated formal verification of object-oriented libraries, and created the first fully verified general-purpose container library, receiving the Best Paper Award at the International Symposium on Formal Methods. During her doctoral studies, Nadia was an intern at MSR Redmond, where she worked on verifying real-world implementations of security protocols. At MIT, Nadia has been applying formal verification to automate various critical and error-prone programming tasks. 32-G449

April 26

EECS Special Seminar: Le Song, "Embedding as a Tool for Algorithm Design"

Le Song
Georgia Tech

Part Of

4:00P
- 5:00P

Location

32-D463
Add to Calendar 2017-04-26 16:00:00 2017-04-26 17:00:00 America/New_York EECS Special Seminar: Le Song, "Embedding as a Tool for Algorithm Design" Abstract: Many big data analytics problems are intrinsically complex and hard, making the design of effective and scalable algorithms very challenging. Domain experts need to perform extensive research, and experiment with many trial-and-errors, in order to craft approximation or heuristic schemes that meet the dual goals of effectiveness and scalability. Very often, restricted assumptions about the data, which are likely to be violated in real world, are made in order for the algorithms to work and obtain performance guarantees. Furthermore, previous algorithm design paradigms seldom systematically exploit a common trait of real-world problems: instances of the same type of problem are solved repeatedly on a regular basis, differing only in their data. Is there a better way to design effective and scalable algorithms for big data analytics?I will present a framework for addressing this challenge based on the idea of embedding algorithm steps into nonlinear spaces, and learn these embedded algorithms from problem instances via either direct supervision or reinforcement learning. In contrast to traditional algorithm design where every steps in an algorithm is prescribed by experts, the embedding design will delegate some difficult algorithm choices to nonlinear learning models so as to avoid either large memory requirement, restricted assumptions on the data, or limited design space exploration. I will illustrate the benefit of this new design framework using large scale real world data, including a materials discovery problem, a recommendation problem over dynamic information networks, and a problem of learning combinatorial algorithms over graphs. The learned algorithms can reduce memory usage and runtime by orders of magnitude, and sometimes result in drastic improvement in predictive performance.Bio: Le Song is an Associate Professor in the Department of Computational Science and Engineering, College of Computing, and an Associate Director of the Center for Machine Learning, Georgia Institute of Technology. He received his Ph.D. in Machine Learning from University of Sydney and NICTA in 2008, and then conducted his post-doctoral research in the Department of Machine Learning, Carnegie Mellon University, between 2008 and 2011. Before he joined Georgia Institute of Technology in 2011, he was a research scientist at Google briefly. His principal research direction is machine learning, especially nonlinear models, such as kernel methods and deep learning, and probabilistic graphical models for large scale and complex problems, arising from artificial intelligence, network analysis, computational biology and other interdisciplinary domains. He is the recipient of the Recsys’16 Deep Learning Workshop Best Paper Award, AISTATS'16 Best Student Paper Award, IPDPS'15 Best Paper Award, NSF CAREER Award’14, NIPS’13 Outstanding Paper Award, and ICML’10 Best Paper Award. He has also served as the area chair or senior program committee for many leading machine learning and AI conferences such as ICML, NIPS, AISTATS, AAAI and IJCAI, and the action editor for JMLR. 32-D463

April 24

EECS Special Seminar: Justin Hsu, "Randomized Algorithms Meets Formal Verification"

Justin Hsu
University of Pennsylania

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-24 16:00:00 2017-04-24 17:00:00 America/New_York EECS Special Seminar: Justin Hsu, "Randomized Algorithms Meets Formal Verification" Abstract: Algorithms and formal verification are two classical areas of computer science. The two fields apply rigorous mathematical proof to seemingly disparate ends---on the one hand, analyzing computational efficiency of algorithms; on the other, designing techniques to mechanically show that programs are correct.In this talk, I will present a surprising confluence of ideas from these two areas. First, I will show how coupling proofs, used to analyze random walks and Markov chains, correspond to proofs in the program logic pRHL (probabilistic Relational Hoare Logic). This connection enables formal verification of novel probabilistic properties, and provides an structured understanding of proofs by coupling. Then, I will show how an approximate version of pRHL, called apRHL, points to a new, approximate version of couplings closely related to differential privacy. The corresponding proof technique---proof by approximate coupling---enables cleaner proofs of differential privacy, both for humans and for formal verification. Finally, I will share some directions towards a possible "Theory AB", blending ideas from both worlds.Bio: Justin Hsu is a final year graduate student in Computer Science at the University of Pennsylvania. He obtained his undergraduate degree in Mathematics from Stanford University. His research interests span formal verification and theoretical computer science, including verification of randomized algorithms, differential privacy, and game theory. He is the recipient of a Simons graduate fellowship in Theoretical Computer Science. 32-G449

April 20

EECS Special Seminar: Arvind Satyanarayanan, "Declarative Interaction Design for Data Visualization"

Arvind Satyanarayanan
Stanford University

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-20 16:00:00 2017-04-20 17:00:00 America/New_York EECS Special Seminar: Arvind Satyanarayanan, "Declarative Interaction Design for Data Visualization" Abstract: Interactive visualization is an increasingly popular medium for analysis and communication as it allows readers to engage data in dialogue. Hypotheses can be rapidly generated and evaluated in situ, facilitating an accretive construction of knowledge and serendipitous discovery. Yet, existing models of visualization relegate interaction to a second-class citizen: imperative event handling callbacks that are difficult to specify, and even harder to reason about.In this talk, I will introduce two new declarative languages that lower the threshold for authoring interactive visualizations, and enable higher-level reasoning about the design space of interactions. Reactive Vega is an expressive representation that is well-suited for custom, explanatory visualizations. It shifts the burden of execution from the user to the underlying streaming dataflow system. Vega-Lite builds on Vega to provide a higher-level grammar for rapidly specifying interactive graphics for exploratory analysis. Its concise format decomposes interaction design into semantic units that can be systematically enumerated.Together, these languages serve as platforms for further research into novel methods of expressing visualization design, and systems for interactive data analysis. And, critically, they provide a growing and engaged community to study their use with -- the Wikipedia and Jupyter communities, for instance, have embraced Vega and Vega-Lite to author interactive visualizations within articles and data science notebooks, respectively.BIO: Arvind Satyanarayan is a Computer Science PhD candidate at Stanford University, working with Jeffrey Heer and the University of Washington Interactive Data Lab. Arvind's research develops new declarative languages for interactive visualization, and leverages them in new systems for visualization design and data analysis. His work has been recognized with a Google PhD Fellowship, Best Paper Awards at IEEE InfoVis & ACM CHI, and has been deployed on Wikipedia to enable interactive visualizations within articles.WEBSITEhttp://arvindsatya.com/ 32-G449

April 19

EECS Special Seminar: Oliver Kroemer, "Learning Robot Manipulation Skills through Experience and Generalization"

Oliver Kroemer
University of Southern California

Part Of

4:00P
- 5:00P

Location

32-D463
Add to Calendar 2017-04-19 16:00:00 2017-04-19 17:00:00 America/New_York EECS Special Seminar: Oliver Kroemer, "Learning Robot Manipulation Skills through Experience and Generalization" Abstract: In the future, robots could be used to take care of the elderly, perform household chores, and assist in hazardous situations. However, such applications require robots to manipulate objects in unstructured and everyday environments. Hence, in order to perform a wide range of tasks, robots will need to learn manipulation skills that generalize between different scenarios and objects.In this talk, Oliver Kroemer will present methods that he has developed for robots to learn versatile manipulation skills. The first part of the presentation will focus on learning manipulation skills through trial and error using task-specific reward functions. The presented methods are used to learn the desired trajectories and goal states of individual skills, as well as high-level policies for sequencing skills. The resulting skill sequences reflect the multi-modal structure of the manipulation tasks.The second part of the talk will focus on representations for generalizing skills between different scenarios. He will explain how robots can learn to adapt skills to the geometric features of the objects being manipulated, and also discuss how the relevance of these features can be predicted from previously learned skills. The talk will be concluded by discussing open challenges and future research directions for learning manipulation skills.Bio: Oliver Kroemer is a postdoctoral researcher at the University of Southern California (USC), working together with Gaurav S. Sukhatme in the Robotic Embedded Systems Lab (RESL). His research interests are in machine learning and robotics, with a focus on learning grasping and manipulation skills. He received his Masters and Bachelors degrees in engineering from the University of Cambridge in 2008. He was a Ph.D. student at the Max Planck Institute for Intelligent Systems from 2009 to 2011. In 2014, Oliver defended his Ph.D. thesis at the Technische Universitaet Darmstadt. He was a finalist for the 2015 Georges Giralt Ph.D. Award for the best robotics Ph.D. thesis in Europe. In 2015, he first worked as a postdoctoral researcher at TU 32-D463

April 13

EECS Special Seminar: David Held, "Robots in Clutter: Learning to Understand Environmental Changes"

David Held
UC Berkeley

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-13 16:00:00 2017-04-13 17:00:00 America/New_York EECS Special Seminar: David Held, "Robots in Clutter: Learning to Understand Environmental Changes" Abstract: Robots today are confined to operate in relatively simple, controlled environments. One reason for this is that current methods for processing visual data tend to break down when faced with occlusions, viewpoint changes, poor lighting, and other challenging but common situations that occur when robots are placed in the real world. I will show that we can train robots to handle these variations by modeling the causes behind visual appearance changes. If robots can learn how the world changes over time, they can be robust to the types of changes that objects often undergo. I demonstrate this idea in the context of autonomous driving, and I will show how we can use this idea to improve performance for every step of the robotic perception pipeline: object segmentation, tracking, and velocity estimation. I will also present some recent work on learning to manipulate objects, using a similar framework of learning environmental changes. By learning how the environment can change over time, we can enable robots to operate in the complex, cluttered environments of our daily lives.Bio: David Held is a post-doctoral researcher at U.C. Berkeley working with Pieter Abbeel on deep reinforcement learning for robotics. He recently completed his Ph.D. in Computer Science at Stanford University with Sebastian Thrun and Silvio Savarese, where he developed methods for perception for autonomous vehicles. David has also worked as an intern on Google’s self-driving car team. Before Stanford, David was a researcher at the Weizmann Institute, where he worked on building a robotic octopus. He received a B.S. and M.S. in Mechanical Engineering at MIT and an M.S. in Computer Science at Stanford, for which he was awarded the Best Master's Thesis Award from the Computer Science Department. 32-G449

April 10

EECS Special Seminar: Julian Shun "Shared-Memory Parallelism Can Be Simple, Fast, and Scalable"

Julian Shun
UC Berkeley

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-10 16:00:00 2017-04-10 17:00:00 America/New_York EECS Special Seminar: Julian Shun "Shared-Memory Parallelism Can Be Simple, Fast, and Scalable" Abstract: Parallelism is the key to achieving high performance in computing. However, writing efficient and scalable parallel programs is notoriously difficult, and often requires significant expertise. To address this challenge, it is crucial to provide programmers with high-level tools to enable them to develop solutions more easily, and at the same time emphasize the theoretical and practical aspects of algorithm design to allow the solutions developed to run efficiently under many possible settings. My research addresses this challenge using a three-pronged approach consisting of the design of shared-memory programming techniques, frameworks, and algorithms for important problems in computing. In this talk, I will present tools for deterministic parallel programming, large-scale shared-memory algorithms that are efficient both in theory and in practice, and Ligra, a framework for simplifying the programming of shared-memory graph algorithms.Bio: Julian Shun is currently a Miller Research Fellow (post-doc) at UC Berkeley. He obtained his Ph.D. in Computer Science from Carnegie Mellon University, and his undergraduate degree in Computer Science from UC Berkeley. He is interested in developing large-scale parallel algorithms for graph processing, and parallel text algorithms and data structures. He is also interested in designing methods for writing deterministic parallel programs and benchmarking parallel programs. He has received the ACM Doctoral Dissertation Award, CMU School of Computer Science Doctoral Dissertation Award, Miller Research Fellowship, Facebook Graduate Fellowship, and a best student paper award at the Data Compression Conference. 32-G449

April 06

EECS Special Seminar: Li-Yang Tan "Computational complexity through the lens of circuits, proofs, and randomness"

Li-Yang Tan
Toyota Technological Institute in Chicago

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-06 16:00:00 2017-04-06 17:00:00 America/New_York EECS Special Seminar: Li-Yang Tan "Computational complexity through the lens of circuits, proofs, and randomness" Abstract: Computational complexity theory is rooted in many of computer science's most fascinating questions. Three examples of such questions are: Are there functions that are simple to describe, and yet require large circuits to compute? Are there short mathematical theorems that require lengthy proofs? Does every efficient randomized algorithm have an efficient deterministic counterpart? In this talk I will describe results motivated by and contributing to our understanding of these questions, focusing on three results: (1) An average-case depth hierarchy theorem for boolean circuits, which resolves a thirty-year-old conjecture in circuit complexity; (2) Exponentially-improved lower bounds for the Frege proof system, the canonical proof system for propositional logic; (3) The fastest deterministic algorithm for finding satisfying assignments of CNF formulas that have many such assignments, a basic unresolved problem in derandomization. These results highlight rich connections among the respective subfields of complexity theory --- circuit complexity, proof complexity, and pseudorandomness --- and also have implications to areas outside of complexity theory, such as parallel computing and SAT solving. I will also touch on how the techniques we have developed point to avenues for progress on a few flagship challenges of complexity theory. Bio: Li-Yang Tan received a PhD in Computer Science from Columbia University in 2014. From 2014 to 2015 he was a Microsoft Research fellow at the UC Berkeley Simons Institute for the Theory of Computing, and since 2015 he has been a research assistant professor at the Toyota Technological Institute in Chicago. His research interests are in computational complexity, with an emphasis on circuit complexity, proof complexity, pseudorandomness, and the analysis of boolean functions. He is a recipient of the Best Paper award at FOCS '15 and an NSF Algorithmic Foundations (Medium) award. 32-G449

April 03

EECS Special Seminar: Veselin Raychev, "Machine Learning for Programming"

Veselin Raychev
ETH, Zurich

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-04-03 16:00:00 2017-04-03 17:00:00 America/New_York EECS Special Seminar: Veselin Raychev, "Machine Learning for Programming" Abstract: In this talk I will discuss a new generation of software tools based on probabilistic models learned from large codebases. These tools leverage the massive effort already spent by thousands of programmers and make useful predictions about new, unseen programs, thus helping to solve important and difficult software tasks. As an example, I will illustrate several such practical systems including statistical code completion, deobfuscation and defect prediction. Two of these systems (jsnice.org and apk-deguard.com) are freely available and already have thousands of users. In addition, I will also present some of the core machine learning techniques underlying our tools. I will discuss new probabilistic models of code that are more precise than state-of-the-art neural networks while requiring fewer computational resources to train and use.Short Bio: Veselin Raychev obtained his PhD from ETH Zurich in 2016 on the topic of “Learning from Large Codebases”. Before this, he worked as a software engineer at Google on the public transportation routing algorithm of Google Maps as well as several other projects. His research interests include machine learning, program analysis, program synthesis and algorithms. 32-G449

EECS Special Seminar: Amanda Prorok, "Diversity and Resilience in Robot Networks"

Amanda Prorok
University of Pennsylvania

Part Of

3:00P
- 4:00P

Location

34-401 (grier)
Add to Calendar 2017-04-03 15:00:00 2017-04-03 16:00:00 America/New_York EECS Special Seminar: Amanda Prorok, "Diversity and Resilience in Robot Networks" Abstract: Recent years have seen falling costs of communication and storage technologies and advances in fabrication methods. Sensors, actuators, and processors are being integrated into globally accessible information networks. These trends are promoting a profusion of networked robotic platforms with distinct features and unique capabilities. As we aspire to harness this diverse array of robots to solve increasingly complex problems, heterogeneity and diversity become design features. However, we still lack a fundamental understanding of how to compose and control large-scale systems of heterogeneous robots. Moreover, as we program diverse robots to exploit their technical complementarities, we create interdependencies and critical links. Such collaborative algorithms give rise to new sources of internal and external threats that lead to unintended failure modes. As a consequence, we need new mechanisms that ensure resilience. I begin my talk by formalizing diversity in the context of dynamic task allocation for large-scale heterogeneous multi-robot systems. In light of this setting, I show how optimal control policies are impacted by the heterogeneity of the robot team. In the second part of the talk, my focus shifts to the question of how to provide resilience to internal failures through precautionary collaboration mechanisms. By building on foundational concepts of network science and security, I show how we can achieve resilience, allowing robot teams to function in the presence of defective and/or malicious robots. Finally, I consider the importance of providing system-wide protection against external threats, and introduce some new ideas that touch upon privacy. Bio: Amanda Prorok is a Postdoctoral Researcher in the General Robotics, Automation, Sensing and Perception (GRASP) Laboratory at the University of Pennsylvania, where she works with Prof. Vijay Kumar on heterogeneous networked robotic systems. She completed her PhD at EPFL, Switzerland, where she addressed the topic of localization with ultra-wideband sensing for robotic networks. Her dissertation was awarded the Asea Brown Boveri (ABB) award for the best thesis at EPFL in the fields of Computer Sciences, Automatics and Telecommunications. She was selected as an MIT Rising Star in 2015, and won a Best Paper Award at the 9th International Conference on Bio-inspired Information and Communications Technologies, 2015. Hosts: Daniela Rus and Gerry Sussman 34-401 (grier)

March 23

EECS Special Seminar: Shivaram Venkataraman, "Scalable Systems for Fast and Easy Machine Learning"

Shivaram Venkataraman
Computer Science at UC Berkeley

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-03-23 16:00:00 2017-03-23 17:00:00 America/New_York EECS Special Seminar: Shivaram Venkataraman, "Scalable Systems for Fast and Easy Machine Learning" Abstract: Machine learning models trained on massive datasets power a number of applications; from machine translation to detecting supernovae in astrophysics. However the end of Moore’s law and the shift towards distributed computing architectures presents many new challenges for building and executing such applications in a scalable fashion. In this talk I will present my research on systems that make it easier to develop new machine learning applications and scale them while achieving high performance. I will first present programming models that let users easily build distributed machine learning applications. Next, I will show how we can simplify large scale deployments and understand scalability with low-overhead performance models. Finally, I will describe scheduling techniques that exploit the structure of machine learning algorithms to improve scalability and achieve high performance when using distributed data processing frameworks.Bio: Shivaram Venkataraman is a PhD Candidate at the University of California, Berkeley and is advised by Mike Franklin and Ion Stoica. His research interests are in designing systems and algorithms for large scale data processing and machine-learning. He is a recipient of the Siebel Scholarship and best-of-conference citations at VLDB and KDD. Before coming to Berkeley, he completed his M.S at the University of Illinois, Urbana-Champaign. 32-G449

March 20

MIT EECS Special Seminar: Jacob Eisenstein, "Computational Sociolinguistics: Social Networks, Social Media, Social Meanings"

Jacob Eisenstein
Georgia Tech

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-03-20 16:00:00 2017-03-20 17:00:00 America/New_York MIT EECS Special Seminar: Jacob Eisenstein, "Computational Sociolinguistics: Social Networks, Social Media, Social Meanings" Abstract: Language is socially situated: both what we say and what we mean are dependent on our social identities, our interlocutors, and the communicative setting. Acknowledging and adapting to social variation will be critical for developing natural language processing systems that are robust across a diverse range of genres, platforms, and authors. In this talk, I will describe research in the nascent field of computational sociolinguistics, which aims to formalize language's social dimension through computational techniques. First, I will show how unsupervised machine learning over social network labelings and text enables the induction of social meanings for address terms, such as "Ms" and "dude". Next, I will describe how to build NLP systems that are robust to social variation. In this work, we use social network embeddings to induce personalized language processing systems for individual social media users, improving performance even for users for whom no labeled data is available. Finally, I will describe how the spread of linguistic innovations can serve as evidence for sociocultural affinity and influence, using a range of computational techniques that include vector autoregressive models, Hawkes processes, and causal inference.Bio: Jacob Eisenstein is an Assistant Professor in the School of Interactive Computing at Georgia Tech. He works on statistical natural language processing, focusing on computational sociolinguistics, social media analysis, discourse, and machine learning. He is a recipient of the NSF CAREER Award, a member of the Air Force Office of Scientific Research (AFOSR) Young Investigator Program, and was a SICSA Distinguished Visiting Fellow at the University of Edinburgh. His work has also been supported by the National Institutes for Health, the National Endowment for the Humanities, and Google. Jacob was a Postdoctoral researcher at Carnegie Mellon and the University of Illinois. He completed his Ph.D. at MIT in 2008, winning the George M. Sprowls dissertation award. Jacob's research has been featured in the New York Times, National Public Radio, and the BBC. Thanks to his brief appearance in If These Knishes Could Talk, Jacob has a Bacon number of 2. 32-G449

March 16

EECS Special Seminar: Luca Carlone, "Convex Relaxations for Lightweight Peception and Sensing on Micro Aerial Vehicles"

Luca Carlone
MIT

Part Of

3:00P
- 4:00P

Location

34-401A
Add to Calendar 2017-03-16 15:00:00 2017-03-16 16:00:00 America/New_York EECS Special Seminar: Luca Carlone, "Convex Relaxations for Lightweight Peception and Sensing on Micro Aerial Vehicles" Abstract: The development of robust and lightweight perception algorithms is crucial towards the deployment of robotics technologies, ranging from self-driving cars to micro aerial vehicles. Perception algorithms are responsible for interpreting sensor data into a coherent world representation, which can be used to support navigation and decision-making. Perception needs to be lightweight, to cope with limited on-board computation; moreover, it needs to produce certifiably correct results, in the face of large measurement noise and outliers.In this talk, I present my work on lightweight and robust robot perception, using agile navigation of micro aerial vehicles as a motivating application. I start by drawing connections between robot perception and optimization, and show that a large class of geometric problems in robotics and computer vision can be cast as a nonconvex optimization problem with variables living on manifold. Then I consider an instance of this nonconvex problem, and present a convex relaxation that is able to recover the exact global solution of the nonconvex problem in a noise regime that encompasses most applications in robotics and computer vision. Besides being certifiably correct and robust to large noise, our convex relaxation is lightweight, entailing a computational cost that is an order of magnitude smaller than standard iterative solvers. After discussing robot perception, I provide a brief overview of our recent work on lightweight sensing. I consider the case in which a small robot does not have sufficient payload to carry a standard depth sensor, and it has to reconstruct the geometry of the environment from a sparse set of noisy depth measurements. Also in this case, I show that the use of convex relaxations, akin the ones used in compressive sensing, enables reconstruction from sparse data. I conclude by proposing few ideas to scale down perception to miniaturized platforms, such as nano and pico aerial vehicles, where sensing and computation are subject to strict payload and power constraints. Bio: Luca Carlone is a research scientist in the Laboratory for Information and Decision Systems at the Massachusetts Institute of Technology. Before joining MIT, he was a postdoctoral fellow at Georgia Tech (2013-2015), and a visiting researcher at the University of California Santa Barbara (2011). He got his Ph.D. from the Polytechnic University of Turin, Italy, in 2012. His research interests include nonlinear estimation, numerical and distributed optimization, computer vision and probabilistic inference applied to sensing, perception, and control of single and multi robot systems. He published more than 60 papers on international journals and conferences, including a best paper award finalist at RSS 2015 and a best paper award winner at WAFR 2016. 34-401A

March 13

EECS Special Seminar: Dorsa Sadigh, "Towards a Theory of Safe and Interactive Autonomy"

Dorsa Sadigh

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-03-13 16:00:00 2017-03-13 17:00:00 America/New_York EECS Special Seminar: Dorsa Sadigh, "Towards a Theory of Safe and Interactive Autonomy" Abstract: Today’s society is rapidly advancing towards cyber-physical systems (CPS) that interact and collaborate with humans, e.g., semi-autonomous vehicles interacting with drivers and pedestrians, medical robots used in collaboration with doctors, or service robots interacting with their users in smart homes. The safety-critical nature of these systems requires us to provide provably correct guarantees about their performance in interaction with humans. The goal of my research is to enable such human-cyber-physical systems (h-CPS) to be safe and interactive. I aim to develop a formalism for design of algorithms and mathematical models that facilitate correct-by-construction control for safe and interactive autonomy.In this talk, I will first discuss interactive autonomy, where we use algorithmic human-robot interaction to be mindful of the effects of autonomous systems on humans, and further leverage these effects for better safety, efficiency, coordination, and estimation. I will then talk about safe autonomy, where we provide correctness guarantees, while taking into account the uncertainty arising from the environment. Further, I will discuss a diagnosis and repair algorithm for systematic transfer of control to the human in unrealizable settings. While the algorithms and techniques introduced can be applied to many h-CPS applications, in this talk, I will focus on the implications of my work for semi-autonomous driving. Bio: Dorsa Sadigh is a Ph.D. candidate in the Electrical Engineering and Computer Sciences department at UC Berkeley. Her research interests lie in the intersection of control theory, formal methods, and human-robot interactions. Specifically, she works on developing a unifying framework for safe and interactive autonomy. Dorsa received her B.S. from Berkeley EECS in 2012. She was awarded the NDSEG and NSF graduate research fellowships in 2013. She was the recipient of the 2016 Leon O. Chua department award and the 2011 Arthur M. Hopkin department award for achievement in the field of nonlinear science, and she received the Google Anita Borg Scholarship in 2016. 32-G449

March 09

EECS Special Seminar: Phillip Isola, "Learning to see without a teacher"

Phillip Isola
UC Berkeley

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-03-09 16:00:00 2017-03-09 17:00:00 America/New_York EECS Special Seminar: Phillip Isola, "Learning to see without a teacher" Abstract: Over the past decade, learning-based methods have driven rapid progress in computer vision. However, most such methods still require a human "teacher" in the loop. Humans provide labeled examples of target behavior, and also define the objective that the learner tries to satisfy. The way learning plays out in nature is rather different: ecological scenarios involve huge quantities of unlabeled data and only a few supervised lessons provided by a teacher (e.g., a parent). I will present two directions toward computer vision algorithms that learn more like ecological agents. The first involves learning from unlabeled data. I will show how objects and semantics can emerge as a natural consequence of predicting raw data, rather than labels. The second is an approach to data prediction where we not only learn to make predictions, but also learn the objective function that scores the predictions. In effect, the algorithm learns not just how to solve a problem, but also what exactly needs to be solved in order to generate realistic outputs. Finally, I will talk about my ongoing efforts toward sensorimotor systems that not only learn from provided data but also act to sample more data on their own. Bio: Phillip Isola is a postdoctoral scholar in the EECS department at UC Berkeley. He recently received his Ph.D. in the Brain & Cognitive Sciences department at MIT. He studies visual intelligence from the perspective of both minds and machines. He was the recipient of both the NSF Graduate Fellowship and, presently, the NSF Postdoctoral Fellowship. 32-G449

March 06

EECS Special Seminar: Tim Kraska "The Unknown Unknowns in Data Integration"

Tim Kraska
Brown University

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-03-06 16:00:00 2017-03-06 17:00:00 America/New_York EECS Special Seminar: Tim Kraska "The Unknown Unknowns in Data Integration" Abstract: It is common practice for data scientists to acquire and integrate disparate data sources into one. Besides the tremendous work on cleaning and integrating techniques, surprisingly, there has been very little work on determining (1) how clean and (2) how complete the data is, and (3) what the potential impact of any unknown data (a.k.a., unknown unknowns) on query results is?In this talk, I will first present new techniques to estimate the completeness and the impact of unknown data on simple aggregate queries. The key idea is that the overlap between different data sources enables us to estimate the number and values of the missing data items. Second, I will present novel techniques to estimate the number of remaining errors in a data set. Finally, I will show how these techniques are integrated in QUDE, a new component of our interactive data exploration system, which aims to automatically assist users in identifying potential risk factors, such as the mentioned missing data items. Bio: Tim Kraska is an Assistant Professor in the Computer Science department at Brown University. Currently, his research focuses building systems for interactive data exploration and transactional systems for modern hardware, especially the next generation of networks. Before joining Brown, Tim spent 3 years as a PostDoc in the AMPLab at UC Berkeley, where he worked on hybrid human-machine database systems and cloud-scale data management systems. Tim received his PhD from the ETH Zurich under the supervision of Donald Kossmann. He was awarded an NSF Career Award (2015), an Airforce Young Investigator award (2015), a Swiss National Science Foundation Prospective Researcher Fellowship (2010), a DAAD Scholarship (2006), a University of Sydney Master of Information Technology Scholarship for outstanding achievement (2005), the University of Sydney Siemens Prize (2005), two VLDB best demo awards (2015 and 2011), and an ICDE best paper award (2013), and very recently got selected as a 2017 Alfred P. Sloan Research Fellow in Computer Science. 32-G449

March 01

EECS Special Seminar: Wyatt Lloyd, "Low Latency and Strong Guarantees for Scalable Storage"

Wyatt Lloyd
Princeton University

Part Of

4:00P
- 5:00P

Location

32-G882
Add to Calendar 2017-03-01 16:00:00 2017-03-01 17:00:00 America/New_York EECS Special Seminar: Wyatt Lloyd, "Low Latency and Strong Guarantees for Scalable Storage" Abstract: Scalable storage systems, where data is sharded across many machines, are necessary to support web services whose data is too large for a single machine to handle. An ideal system would provide the lowest latency—to make the web services built on top of it fast—and the strongest guarantees—to make programming the web service easier. Theoretical results prove that such an ideal system is impossible, but all hope is not lost! Our work has made progress on this problem from both directions: providing stronger guarantees for low latency systems and providing lower latency for systems with strong guarantees. I will cover one of these systems, Eiger, in detail. I will also touch on our recent impossibility result, the SNOW Theorem, and how it guided us in building new systems with latency-optimal read-only transactions.Bio: Wyatt Lloyd is a third-year assistant professor of computer science at the University of Southern California. His research interests include all aspects of large-scale distributed systems. He received his Ph.D. from Princeton University in 2013. He then spent a year as a Postdoctoral Researcher at Facebook, and he continues to collaborate with its engineers on projects related to media processing, storage, and delivery. 32-G882

February 27

EECS Special Seminar: Irene Zhang, "Towards New Systems for Mobile/Cloud Applications"

Irene Zhang
University of Washington

Part Of

4:00P
- 5:00P

Location

32-G449
Add to Calendar 2017-02-27 16:00:00 2017-02-27 17:00:00 America/New_York EECS Special Seminar: Irene Zhang, "Towards New Systems for Mobile/Cloud Applications" Abstract: The proliferation of datacenters, smartphones, personal sensing andtracking devices, and home automation products is fundamentallychanging the applications we interact with daily. Today's popularuser applications are no longer limited to a single desktop computerbut now commonly span many mobile devices and cloud servers. As aresult, existing systems often do not meet the needs of modernmobile/cloud applications. In this talk, I will present three systemsdesigned to tackle the challenges of mobile/cloud applications:Sapphire, Diamond and TAPIR. These systems represent a first steptowards better end-to-end support for mobile/cloud applications inruntime management, data management, and distributed transactionalstorage. Together, they significantly simplify the development of newmobile/cloud applications.Bio: Irene Zhang is a fifth-year PhD student at the University ofWashington. She works with Hank Levy and Arvind Krishnamurthy in theComputer Systems Lab. Her current research focuses on systems forlarge-scale, distributed applications, including distributed runtimesystems and transactional storage systems. Before starting her PhD,Irene worked for three years at VMware in the virtual machine monitorgroup. Irene received her S.B. and M. Eng. in computer science fromMIT, where she worked with Frans Kaashoek in the Parallel andDistributed Operating Systems Lab. 32-G449

February 23

EECS Special Seminar: Tengyu Ma, "Better understanding of non-convex methods in machine learning"

Tengyu Ma
Princeton University

Part Of

3:00P
- 4:00P

Location

34-401A
Add to Calendar 2017-02-23 15:00:00 2017-02-23 16:00:00 America/New_York EECS Special Seminar: Tengyu Ma, "Better understanding of non-convex methods in machine learning" Abstract: Recent breakthroughs in machine learning, especially deep learning, often involve learning complex and high-dimensional models on massive datasets with non-convex optimization. Though empirically successful, the formal study of such non-convex methods is much less developed. My research aims to develop new algorithmic approaches and analysis tools in these settings. The talk will showcase a few results. First, we show that matrix completion — a famous problem in machine learning — can be solved by stochastic gradient descent on the straightforward non-convex objective function in polynomial time. (Formally, we show that all local minima of the objective are also global minima.) Then, we will analyze the landscape of the objective functions for linearized recurrent neural nets and residual nets, and demonstrate that over-parameterization and re-parameterization of the models can make the optimization easier. Bio: Tengyu Ma is a PhD candidate at the Computer Science Department of Princeton University advised by Sanjeev Arora. His research interests include topics in machine learning and algorithms, such as non-convex optimization, representation learning, deep learning, and convex relaxation for machine learning problems. His research contributes to the theoretical developments of these topics with practical implication. He is a recipient of NIPS'16 best student paper award, Princeton Honorific Fellowship, Siebel Scholarship, IBM PhD Fellowship, and the Simons Award for Graduate Students in Theoretical Computer Science. 34-401A

February 21

EECS/IDSS Special Seminar: Justin Cheng, "Antisocial Computing: Explaining and Predicting Negative Behavior Online"

Justin Cheng
Stanford University

Part Of

4:00P
- 5:00P

Location

32-D463
Add to Calendar 2017-02-21 16:00:00 2017-02-21 17:00:00 America/New_York EECS/IDSS Special Seminar: Justin Cheng, "Antisocial Computing: Explaining and Predicting Negative Behavior Online" Abstract: Antisocial behavior and misinformation are increasingly prevalent online. As usersinteract with one another on social platforms, negative interactions can cascade, resulting incomplex changes in behavior that are difficult to predict. My researchintroduces computational methods for explaining the causes of such negative behavior and forpredicting its spread in online communities. It complements data mining with crowdsourcing,which enables both large-scale analysis that is ecologically valid and experiments that establishcausality. First, in contrast to past literature which has characterized trolling as confined to avocal, antisocial minority, I instead demonstrate that ordinary individuals, under the rightcircumstances, can become trolls, and that this behavior can percolate and escalate througha community. Second, despite prior work arguing that such behavioral and informationalcascades are fundamentally unpredictable, I demonstrate how their future growth can bereliably predicted. Through revealing the mechanisms of antisocial behavior online, my workexplores a future where systems can better mediate interpersonal interactions and insteadpromote the spread of positive norms in communities.Bio: Justin Cheng is a PhD candidate in the Computer Science Department at StanfordUniversity, where he is advised by Jure Leskovec and Michael Bernstein. His research lies at theintersection of data science and human-computer interaction, and focuses on cascadingbehavior in social networks. This work has received a best paper award, as well as several bestpaper nominations at CHI, CSCW, and ICWSM. He is also a recipient of a Microsoft ResearchPhD Fellowship and a Stanford Graduate Fellowship. 32-D463