Graduate

EE380 Computer Systems Colloquium: Computer Accessibility

Topic: 
Exploring the implications of machine learning for people with cognitive disabilities
Abstract / Description: 

Advances in information technology have provided many benefits for people with disabilities, including wide availability of textual content via text to speech, flexible control of motor wheelchairs, captioned video, and much more. People with cognitive disabilities benefit from easier communication, and better tools for scheduling and reminders. Will advances in machine learning enhance this impact? Progress in natural language processing, autonomous vehicles, and emotion detection, all driven by machine learning, may deliver important benefits soon. Further out, can we look for systems that can help people with cognitive challenges understand our complex world more easily, work more effectively, stay safe, and interact more comfortably in social situations? What are the technical barriers to overcome in pursuing these goals, and what are the theoretical developments in machine learning that may overcome them?

Date and Time: 
Wednesday, April 18, 2018 - 4:30pm
Venue: 
Gates B03

EE380 Computer Systems Colloquium: Information Theory of Deep Learning

Topic: 
Information Theory of Deep Learning
Abstract / Description: 

I will present a novel comprehensive theory of large scale learning with Deep Neural Networks, based on the correspondence between Deep Learning and the Information Bottleneck framework. The new theory has the following components:

  1. rethinking Learning theory; I will prove a new generalization bound, the input-compression bound, which shows that compression of the representation of input variable is far more important for good generalization than the dimension of the network hypothesis class, an ill defined notion for deep learning.
  2. I will prove that for large scale Deep Neural Networks the mutual information on the input and the output variables, for the last hidden layer, provide a complete characterization of the sample complexity and accuracy of the network. This makes the information Bottleneck bound for the problem as the optimal trade-off between sample complexity and accuracy with ANY learning algorithm.
  3. I will show how Stochastic Gradient Descent, as used in Deep Learning, achieves this optimal bound. In that sense, Deep Learning is a method for solving the Information Bottleneck problem for large scale supervised learning problems. The theory provide a new computational understating of the benefit of the hidden layers, and gives concrete predictions for the structure of the layers of Deep Neural Networks and their design principles. These turn out to depend solely on the joint distribution of the input and output and on the sample size.

Based partly on works with Ravid Shwartz-Ziv and Noga Zaslavsky.

Date and Time: 
Wednesday, April 4, 2018 - 4:30pm
Venue: 
Gates B03

SmartGrid Seminar: Transmission-Distribution Coordinated Energy Management: A Solution to the Challenge of Distributed Energy Resource Integration

Topic: 
Transmission-Distribution Coordinated Energy Management: A Solution to the Challenge of Distributed Energy Resource Integration
Abstract / Description: 

Transmission-distribution coordinated energy management (TDCEM) is recognized as a promising solution to the challenge of high DER penetration, but lack of a distributed computation method that universally and effectively works for TDCEM. To bridge this gap, a generalized master-slave-splitting (G-MSS) method is proposed based on a general-purpose transmission-distribution coordination model (G-TDCM), enabling G-MSS to be applicable to most central functions of TDCEM. In G-MSS, a basic heterogeneous decomposition (HGD) algorithm is first derived from the heterogeneous decomposition of the coupling constraints in the KKT system regarding G-TDCM. Optimality and convergence properties of this algorithm are proved. Furthermore, a modified HGD algorithm is developed by utilizing subsystem's response function, resulting in faster convergence. The distributed G-MSS method is then demonstrated to successfully solve central functions of TDCEM including power flow, contingency analysis, voltage stability assessment, economic dispatch and optimal power flow. Severe issues of over-voltage and erroneous assessment of the system security that are caused by DERs are thus resolved by G-MSS with modest computation cost. A real-world demonstration project in China will be presented.

Date and Time: 
Thursday, April 5, 2018 - 1:30pm
Venue: 
Y2E2 111

US-ATMC (EE402) Seminar presents Asia Entrepreneurship Update – 2018

Topic: 
Asia Entrepreneurship Update – 2018
Abstract / Description: 

In this first session in our Spring Quarter weekly series on "Entrepreneurship in Asian High-Tech Industries," Professor Richard Dasher introduces new (updated) data and discusses trends in entrepreneurism and the supporting ecosystems for startup companies in major Asian economies, as well as the implications for U.S. investors and businesses.

Date and Time: 
Tuesday, April 3, 2018 - 4:30pm
Venue: 
Skilling Auditorium, 494 Lomita Mall

Special Seminar: New Algorithms and Hardware Acceleration for the IoT Revolution

Topic: 
New Algorithms and Hardware Acceleration for the IoT Revolution
Abstract / Description: 

Deep Neural Networks (DNNs) are computation intensive. Without efficient hardware implementation of DNNs, many promising IoT (Internet of Things) applications will not be practically realizable. In this talk, we will first take a detailed look into one type of compute accelerators, FPGA, and evaluate its potential role in the upcoming IoT revolution. Although FPGAs can provide desirable customized hardware solutions, they are difficult to program and optimize. High-level synthesis is an effective design flow for FPGAs due to improved productivity, debugging, and design space exploration abilities. However, optimizing large DNNs under resource constraints for FPGAs is still a key challenge. We will present a series of effective design techniques for implementing DNNs on FPGAs with high performance and energy efficiency. These include the use of configurable DNN IPs, resource allocation across DNN layers, Winograd and FFT techniques, and DNN reduction and re-training. We showcase several design solutions including Long-term Recurrent Convolution Network (LRCN) for video captioning, Inception module (GoogleNet) for face recognition, as well as Long Short-Term Memory (LSTM) for sound recognition. We will also present some of our recent work on developing new DNN models and data structures for achieving higher accuracy for several interesting applications such as crowd counting, genomics, and music modeling.

Date and Time: 
Monday, April 9, 2018 - 4:00pm
Venue: 
Gates 104

SystemX Seminar: Modeling and Simulation for neuromorphic applications with focus on RRAM and ferroelectric devices

Topic: 
Modeling and Simulation for neuromorphic applications with focus on RRAM and ferroelectric devices
Abstract / Description: 

Neuromorphic computing has recently emerged as one of the most promising option to reduce power consumption of big data analysis, paving the way for artificial intelligence systems with power efficiencies like the human brain. The key device for neuromorphic computing system is given by artificial two-terminal synapses controlling signal processing and transmission. Their conductivity must be changed in an analog/continuous way depending on neural signal strengths. In addition, synaptic devices must have: symmetric/linear conductivity potentiation and depression; a high number of levels (~32), which depend on applications and algorithm performances; high data retention (>10 years) and cycling (>109); ultra-low power consumption (<10fJ); low variability; high scalability (<10nm) and possibility of 3D integration.

A variety of different device technologies have been explored such as phase change memories, ferroelectric random-access memory and resistive random-access memory (RRAM). In each case matching the desired specs is a complex multivariable problem requiring a deep quantitative understanding of the link between material properties at the atomic scale and electrical device performance. We have used a multiscale modeling platform GINESTRATM to illustrate this for the case of RRAM and Ferroelectric tunnel junctions (FTJ).

In the case of RRAM, modeling of key mechanisms shows that a dielectric stack composed of two appropriately chosen dielectrics provides the best solution, in agreement with experimental data. In the case of FTJ, the hysteretic ferroelectric behavior of dielectric stacks fabricated from the orthorhombic phase of doped HfO2 is nicely captured by the simulations. These show that Fe-HfO2 stack can be easily used for analog switching by simply tuning set/reset voltage amplitudes. An added advantage of the simulations is that they point out ways to improve the performance, variability and endurance of the devices in order to meet industrial requirements.

Date and Time: 
Thursday, April 5, 2018 - 4:30pm
Venue: 
Gates B03

Special Seminar: Synthesis and Properties of Single-walled Carbon Nanotubes Wrapped with Mono- and Few-layer BN Tubes

Topic: 
Synthesis and Properties of Single-walled Carbon Nanotubes Wrapped with Mono- and Few-layer BN Tubes
Abstract / Description: 

We propose a conceptually new structure, in which mono- or few BN layers seamlessly wrap around a single-walled carbon nanotube (SWNT), and result in an atomically smooth coaxial tube consisting two different materials, as shown in Figure 1. The structure is synthesized by chemical vapor deposition (CVD). As the reaction occurs on outer surface of the existing SWNTs, we name this process conformal CVD. Various SWNTs, e.g. vertically aligned array, horizontally aligned arrays, suspended SWNTs, random network and films, are employed as the starting material, and successful coating are achieved on all of them. Our characterizations confirm that the outside BN coating started locally on the wall of a SWNT and then merge into a BN nanotube on the curved surface of the SWNT which served as a template. The number of walls can be tuned from 1 to few by controlling the CVD condition. The structure of inside SWNTs are almost not effected by the conformal CVD, as evidenced by Raman and many other characterizations. The crystallization and cleanness of the starting SWNT template are believed to be critical for the successful fabrication of outside walls. This structure is expected to have a broad interest and impact in many fields, which include but not limited in investigating the intrinsic optical properties of environment-isolated SWNTs, fabricating BN-protected or gated SWNT devices, and building more sophisticated 1D material systems.

Part of this work was supported by JSPS KAKENHI Grant Numbers JP25107002 and JP15H05760.

Date and Time: 
Thursday, April 5, 2018 - 3:00pm
Venue: 
MERL (Bldg 660), Room 203

Special Seminar: Analog Cybersecurity and Transduction Attacks

Topic: 
Analog Cybersecurity and Transduction Attacks
Abstract / Description: 

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 also suggest educational opportunities for embedded security and discuss how to design out analog cybersecurity risks by rethinking the computing stack from electrons to bits. This work brings some closure to my curiosity on why my cordless phone would ring whenever I executed certain memory operations on the video graphics chip of an Apple IIGS.

Date and Time: 
Tuesday, April 3, 2018 - 3:00pm
Venue: 
Gates 358

SCIEN presents Video-based Reconstruction of the Real World in Motion

Topic: 
Video-based Reconstruction of the Real World in Motion
Abstract / Description: 

New methods for capturing highly detailed models of moving real world scenes with cameras, i.e., models of detailed deforming geometry, appearance or even material properties, become more and more important in many application areas. They are needed in visual content creation, for instance in visual effects, where they are needed to build highly realistic models of virtual human actors. Further on, efficient, reliable and highly accurate dynamic scene reconstruction is nowadays an important prerequisite for many other application domains, such as: human-computer and human-robot interaction, autonomous robotics and autonomous driving, virtual and augmented reality, 3D and free-viewpoint TV, immersive telepresence, and even video editing.

The development of dynamic scene reconstruction methods has been a long standing challenge in computer graphics and computer vision. Recently, the field has seen important progress. New methods were developed that capture - without markers or scene instrumentation - rather detailed models of individual moving humans or general deforming surfaces from video recordings, and capture even simple models of appearance and lighting. However, despite this recent progress, the field is still at an early stage, and current technology is still starkly constrained in many ways. Many of today's state-of-the-art methods are still niche solutions that are designed to work under very constrained conditions, for instance: only in controlled studios, with many cameras, for very specific object types, for very simple types of motion and deformation, or at processing speeds far from real-time.

In this talk, I will present some of our recent works on detailed marker-less dynamic scene reconstruction and performance capture in which we advanced the state of the art in several ways. For instance, I will briefly show new methods for marker-less capture of the full body (like our VNECT approach) and hands that work in more general environments, and even in real-time and with one camera. I will then show some of our work on high-quality face performance capture and face reenactment. Here, I will also illustrate the benefits of both model-based and learning-based approaches and show how different ways to join the forces of the two open up new possibilities. Live demos included!

Date and Time: 
Wednesday, March 21, 2018 - 4:30pm
Venue: 
Packard 101

Reynolds Memorial Seminar presents "Toward Pervasive Robots"

Topic: 
Toward Pervasive Robots
Abstract / Description: 

The digitization of practically everything coupled with the mobile Internet, the automation of knowledge work, and advanced robotics promises a future with democratized use of machines and wide-spread use of robots and customization. However, pervasive use of robots remains a hard problem. For any give task, the body of the robot has to be able to execute the task and the brain of the robot has to be able to control the body to deliver on that task. How can we accelerate the creation of robots customized to specific tasks? Where are the gaps that we need to address in order to advance toward a future where robots are common in the world and they help reliably with physical tasks? What are the roles of design, fabrication, and control along this trajectory?

In this talk Prof. Rus will discuss the use of computational design and fabrication toward pervasive use of robots. Prof. Rus will introduce recent developments in algorithms for customizing robots, focusing on a suite of algorithms for automatically designing, fabricating, and tasking robots using modularity, soft materials, and print-and-fold approaches. She will also describe how computation can play a role in creating robots more capable of reasoning in the world. By enabling on-demand creation of robots, we can begin to imagine a world with one robot for every physical task.

Date and Time: 
Thursday, March 15, 2018 - 5:30pm
Venue: 
Building 530, Room 127

Pages

Subscribe to RSS - Graduate