Back to Events

Seminar Series

June 18

Connections between Circuit Analysis Problems and Circuit Lower Bounds

Cody Murray
MIT CSAIL

Part Of

3:00P
- 4:00P

Location

32-D463 (Star)
Add to Calendar 2018-06-18 15:00:00 2018-06-18 16:00:00 America/New_York Connections between Circuit Analysis Problems and Circuit Lower Bounds Abstract: A circuit analysis problem takes a Boolean function f as input (where f is represented either as a logical circuit, or as a truth table) and determines some interesting property of f. Examples of circuit analysis problems include Circuit Satisfiability, Circuit Composition, and the Minimum Size Circuit Problem (MCSP). A circuit lower bound presents an interesting function f and shows that no "easy" family of logical circuits can compute f correctly on all inputs, for some definition of "easy". Lower bounds are infamously hard to prove, but are of significant interest for understanding computation.In this thesis, we derive new connections between circuit analysis problems and circuit lower bounds, to prove new lower bounds for various well-studied circuit classes. We show how faster algorithms for Circuit Satisfiability can imply non-uniform lower bounds for functions in NP and related classes. We prove that MCSP cannot be NP-hard under "local" gadget reductions of the kind that appear in textbooks, and if MCSP proved to be NP-hard in the usual (polynomial-time reduction) sense then we would also prove longstanding lower bounds in other regimes. We also prove that natural versions of the Circuit Composition problem do not have small circuits that are constructible in very small (logarithmic) space. 32-D463 (Star)

High-resolution Tactile Sensing for Robotic Perception

Wenzhen Yuan
Department of Mechanical Engineering, MIT

Part Of

10:00A
- 11:00A

Location

32-G449
Add to Calendar 2018-06-18 10:00:00 2018-06-18 11:00:00 America/New_York High-resolution Tactile Sensing for Robotic Perception Abstract:Why is it so difficult for the present-day robots to act intelligently in the real-world environment? A major challenge lies in the lack of adequate tactile sensing technologies. Robots need tactile sensing to understand the physical environment, and detect the contact states during manipulation. A recently developed high-resolution tactile sensor, GelSight, which measures detailed information about the geometry and traction field on the contact surface, shows substantial potential for extending the application of robot tactile sensing. The major questions are: (1) What physical information is available from the high-resolution sensor? (2) How can the robot interpret and use this information?This thesis aims at addressing the two questions above. On the one hand, the tactile feedback helps robots on manipulation tasks. I investigate various techniques for detecting incipient slip and full slip during contact with objects. On the other hand, tactile sensing also helps a robot to better understand the physical environment, i.e. exploration tasks. I will present my work on using tactile sensing to estimate the hardness of arbitrary objects, and making a robot autonomously explore the comprehensive properties of common clothing. I also show our work on unsupervised exploration of latent properties of fabrics through cross-modal learning with vision and touch. 32-G449

May 21

Modern Challenges in Distribution Testing

Gautam Kamath
CSAIL MIT

Part Of

11:00A
- 12:00P

Location

32-D463 (Star)
Add to Calendar 2018-05-21 11:00:00 2018-05-21 12:00:00 America/New_York Modern Challenges in Distribution Testing Abstract: Given samples from an unknown distribution p, does it satisfy some hypothesis? This question has received enormous attention in statistics, information theory, and theoretical computer science. Classically, hypothesis testing has been studied in the asymptotic setting. However, a flurry of recent work has focused on non-asymptotic setting, and determining the sample complexity required to achieve non-trivial error rates. In my thesis, I develop methods for hypothesis testing in a number of settings of modern interest, with a focus on sample and computational efficiency. Some settings considered include testing against composite hypotheses, testing with multivariate data, testing on sensitive data, and testing when we have stronger access to the underlying distribution. I will discuss our results as well as interesting directions for future study. 32-D463 (Star)

May 18

"Building and Controlling Fluidically Actuated Soft Robots: From Open Loop to Model-based Control"

Robert Kevin Katzschmann
CSAIL

Part Of

2:30P
- 3:30P

Location

Seminar Room G449 (Patil/Kiva)
32-G449 (Patil/Kiva)
Add to Calendar 2018-05-18 14:30:00 2018-05-18 15:30:00 America/New_York "Building and Controlling Fluidically Actuated Soft Robots: From Open Loop to Model-based Control" Thesis title: "Building and Controlling Fluidically Actuated Soft Robots: From Open Loop to Model-based Control"Author: Robert Kevin KatzschmannDate: May 18th, 2018Time: 2:30 P.M.Location: 32-G449 Patil/Kiva, Stata CenterAbstract:This thesis describes the creation and control of soft robots made of deformable elastomer materials and powered by fluidics. We embed soft fluidic actuators into self-contained soft robotic systems, such as fish for underwater exploration or soft arms for dynamic manipulation. We present models describing the physical characteristics of these continuously deformable and fully soft robots, and then leverage these models for motion planning and closed-loop feedback control in order to realize quasi-static manipulation, dynamic arm motions, and dynamic interactions with an environment.The design and fabrication techniques for our soft robots include the development of soft actuator morphologies, soft casting techniques, as well as closed-circuit pneumatic and hydraulic powering methods. With a modular design approach, we combine these soft actuator morphologies into robotic systems. We create a robotic fish for underwater locomotion, as well as multi-finger hands and multi-segment arms for the use in object manipulation and interaction with an environment. The robotic fish uses a soft hydraulic actuator as its deformable tail to perform open-loop controlled swimming motions through cyclic undulation. The swimming movement of the robotic fish is enabled by a custom-made displacement pump and a custom-made buoyancy control unit, all embedded within a fish-like underwater system. The fish robot receives high-level control commands via acoustic signals to move in marine environments. The control of the multi-segment arms is enabled by models describing the geometry, kinematics, impedance, and dynamics. We use the models for quasi-static closed-loop control and dynamic closed-loop control of the soft arms. The quasi-static controllers work in combination with the kinematic models and geometric motion planners to enable the soft arms to move in confined spaces, and to autonomously perform object grasping. Leveraging the models for impedance and dynamics, we also demonstrate dynamic arm motions and end-effector interactions of the arm with an environment. Our dynamic model allows the application of control techniques developed for rigid robots to be applied for the dynamic control of soft robots. The resulting model-based closed-loop controllers enable dynamic curvature tracking as well as surface following in Cartesian space.Committee:Daniela Rus (Advisor)Andrew and Erna Viterbi Professor of EECSEmail: rus@csail.mit.eduRuss Tedrake (Chair)Toyota Professor of EECS, Aero/Astro, MechE Email: russt@mit.eduAnette HosoiNeil and Jane Pappalardo Professor of Mechanical Engineering Email: peko@mit.eduJohn LeonardSamuel C. Collins Professor of Mechanical and Ocean Engineering Email: jleonard@mit.edu Seminar Room G449 (Patil/Kiva)

May 17

Why Do Approximate Algorithms Work Well in Machine Learning?

Ludwig Schmidt
CSAIL & EECS

Part Of

3:00P
- 4:00P

Location

32-D463 (Star)
Add to Calendar 2018-05-17 15:00:00 2018-05-17 16:00:00 America/New_York Why Do Approximate Algorithms Work Well in Machine Learning? Abstract: Many success stories in machine learning share an intriguing algorithmic phenomenon: while the core algorithmic problems might seem costly to solve or even intractable at first, simple heuristics or approximation algorithms often perform surprisingly well in practice. Common examples include optimizing over non-convex functions or non-convex sets. Even in convex problems, we often settle for sub-optimal solutions returned by stochastic gradient descent. And in nearest neighbor search, a variety of algorithms works remarkably well considering the “curse of dimensionality”.In this thesis, we study this phenomenon in the context of three algorithmic problems that appear widely in the data sciences: constrained optimization (what non-convex sets can we optimize over?), unconstrained optimization (why don’t we compute high-accuracy ERM solutions?), and nearest neighbor search (can the LSH framework explain the empirical state of the art?). The common theme is that the computational hardness of many algorithmic problems appears only below the inherent noise floor of the overall statistical problem. 32-D463 (Star)

Thesis Defense: Large-Scale Probabilistic Aerial Reconstruction

Randi Cabezas
MIT CSAIL

Part Of

1:00P
- 2:00P

Location

32-G449 (Kiva)
Add to Calendar 2018-05-17 13:00:00 2018-05-17 14:00:00 America/New_York Thesis Defense: Large-Scale Probabilistic Aerial Reconstruction Abstract: While much emphasis has been placed on large-scale 3D scene reconstruction from a single data source such as images or distance sensors, models that jointly utilize multiple data types remain largely unexplored. In this work, we will present a Bayesian formulation of scene reconstruction from multi-modal data as well as two critical components that enable large-scale reconstructions with adaptive resolution and high-level scene understanding with meaningful prior-probability distributions.Our first contribution is to formulate the 3D reconstruction problem within the Bayesian framework. We develop an integrated probabilistic model that allows us to naturally represent uncertainty and to fuse complementary information provided by different sensor modalities (imagery and LiDAR). Maximum-a-Posteriori inference within this model leverages GPGPUs for efficient likelihood evaluations. Our dense reconstructions (triangular mesh with texture information) are feasible with fewer observations of a given modality by relaying on others without sacrificing quality.Secondly, to enable large-scale reconstructions our formulation supports adaptive resolutions in both appearance and geometry. This change is motivated by the need for a representation that can adjust to a wide variability in data quality and availability. By coupling edge transformations within a reversible-jump MCMC framework, we allow changes in the number of triangles and mesh connectivity. We demonstrate that these data-driven updates lead to more accurate representations while reducing modeling assumptions and utilizing fewer triangles.Lastly, to enable high-level scene understanding, we include a categorization of reconstruction elements in our formulation. This scene-specific classification of triangles is estimated from semantic annotations (which are noisy and incomplete) and other scene features (e.g., geometry and appearance). The categorization provides a class-specific prior-probability distribution, thus helping to obtain more accurate and interpretable representations by regularizing the reconstruction. Collectively, these models enable complex reasoning about urban scenes by fusing all available data across modalities, a crucial necessity for future autonomous agents and large-scale augmented-reality applications.Committee: John W. Fisher, Polina Golland and John J. Leonard 32-G449 (Kiva)

May 16

Protecting User Data in Large-Scale Web Services

Frank Wang
MIT CSAIL

Part Of

3:30P
- 4:30P

Location

32-G882
Add to Calendar 2018-05-16 15:30:00 2018-05-16 16:30:00 America/New_York Protecting User Data in Large-Scale Web Services Abstract: Web services like Google, Facebook, and Dropbox are now an essential part of people’s lives. In order to provide value to users, these services collect, store, and analyze large amounts of their users’ sensitive data. However, once the user provides her information to the web service, she loses control over how the application manipulates that data. For example, a user cannot control where the application forwards her data. Even if the service wanted to allow users to define access controls, it is unclear how these access controls should be expressed and enforced. Not only is it difficult to develop these secure access control mechanisms, but it is also difficult to ensure these mechanisms are practical. My research addresses these concerns. More specifically, it focuses on building practical, secure mechanisms for protecting user data in large-scale, distributed web services.Bio: Frank Wang is a Ph.D. student at the MIT CSAIL, advised by Nickolai Zeldovich and James Mickens. He completed his undergraduate studies at Stanford University, focusing on applied cryptography. He runs the MIT security seminar and co-founded a summer program for early stage security companies called Cybersecurity Factory. Committee: Nickolai Zeldovich, James Mickens, Vinod Vaikuntanathan 32-G882

Justin Holmgren: Securing Computation on Untrusted Platforms

Justin Holmgren

Part Of

12:30P
- 2:30P

Location

Patil/Kiva G449
Add to Calendar 2018-05-16 12:30:00 2018-05-16 14:30:00 America/New_York Justin Holmgren: Securing Computation on Untrusted Platforms Abstract:In today's networked world, weak devices increasingly rely on remote servers both to store data and to perform costly computations. Unfortunately, these servers may be easily hackable or otherwise untrustworthy. Therefore, without assuming honest behavior on the server's part, we would like to guarantee two basic security objectives:1. (Correctness) It is possible to verify the correctness of the server's computations much more efficiently than by re-executing the computation.2. (Privacy) A server learns nothing about the computation it performs, other than (perhaps) the output.I will present recent results that achieve both these goals for arbitrary computations, and I will conclude with a discussion of open problems and future directions.Thesis Committee: Ran Canetti, Shafi Goldwasser and Vinod Vaikuntanathan Patil/Kiva G449

May 14

Improving Clinical Decisions Using Correspondence Within and Across Electronic Health Records

Jen Gong
MIT - CSAIL

Part Of

12:00P
- 1:00P

Location

Star (32-D463)
Add to Calendar 2018-05-14 12:00:00 2018-05-14 13:00:00 America/New_York Improving Clinical Decisions Using Correspondence Within and Across Electronic Health Records Abstract:Electronic Health Record (EHR) adoption and large-scale retrospective analyses of health care data are part of a broader conversation about health care quality and cost in the United States. Clinical decision-making aids are one method of helping to improve quality and lower cost of care. In this thesis, we present three methods of leveraging correspondences across elements in health care records to aid clinicians in making care decisions. We focus on the critical care environment, where patient state can rapidly change and many care decisions need to be made in short periods of time.First, we introduce a method to leverage correspondences between structured fields from two different EHR systems to a shared space of clinical concepts encoded in an existing domain ontology. We use these correspondences to enable the transfer of machine learning models across different or evolving EHR systems. Second, we introduce a method to learn correspondences between structured health record data and topic distributions of clinical notes written by care team members. Finally, we present a method to characterize care processes by learning correspondences between observations of patient state and actions taken by care team members.Bio:Jen Gong is a Ph.D. candidate in the Data Driven Inference Group at MIT, supervised by John Guttag. Her research focuses on the application of machine learning to healthcare. She is interested in how different modalities of health care data (e.g., structured health record data, clinical notes, physiological time-series) and auxiliary sources (e.g., data from similar patient populations, expert-encoded ontologies) can be leveraged to improve clinical decision-making aids. Prior to MIT, Jen received an A.B. in Applied Mathematics from Harvard College.Committee: John Guttag, Collin Stultz, Jenna Wiens Star (32-D463)

May 07

Computational Design for the Next Manufacturing Revolution

Adriana Schulz
MIT CSAIL Computational Fabrication Group

Part Of

1:00P
- 3:00P

Location

32-D463 (Stata Center - Star Conference Room)
Add to Calendar 2018-05-07 13:00:00 2018-05-07 15:00:00 America/New_York Computational Design for the Next Manufacturing Revolution Abstract: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. Thesis Advisor: Wojciech MatusikThesis Committee: Daniela Rus and Eitan Grinspun 32-D463 (Stata Center - Star Conference Room)