Intelligent High-performance Computing: Classical and Quantum
Tsung-Wei Huang,
University of Utah
When:
11 a.m.–12 p.m., Friday, February 17
Where:
435 Min H. Kao Building
Abstract:
Heterogeneous computing has advanced our application performance to a new level that is previously out of reach. For example, with CPU-GPU heterogeneous parallelism, training a neural network can be 100x faster than CPUs alone. However, writing a program that utilizes heterogeneous computing resources is not an easy job because we need to manage a lot of technical details, such as programming abstraction, scheduling, and concurrency control.
As a result, this talk will introduce a new task-parallel programming system we established to streamline the building of high-performance computing applications. I will talk about key innovations in our system, including a novel programming model and a learning-based runtime, that offer programmer productivity and performance portability. Finally, I will demonstrate how we applied our system to accelerate both classical and quantum computing problems. So far, our system has been downloaded million times, being used by many academic and industrial projects.
Biography:
Tsung-Wei Huang is an assistant professor in the ECE Department at the University of Utah. He received his PhD from the ECE Department at the University of Illinois at Urbana-Champaign and his BS/MS from the CS Department at Taiwan’s National Cheng Kung University. His research group has been creating software systems to streamline the building of high-performance computing applications, including machine learning, computer-aided design, and quantum computing. Huang has received several awards for his research contributions, including ACM SIGDA Outstanding PhD Dissertation Award, NSF CAREER Award, and Humboldt Research Fellowship Award.
Light in Artificial Intelligence: Hardware/Software Co-Design for Photonic Machine Learning Computing
Jiaqi Gu,
The University of Texas at Austin
When:
11 a.m.–12 p.m., Tuesday, February 21
Where:
435 Min H. Kao Building
Abstract:
In the post-Moore era, conventional electronic digital computers have become a limiting factor in certain domains, most notably intelligent information processing. The proliferation of big data and artificial intelligence (AI) has motivated the investigation of next-generation AI computing hardware to support massively parallel and energy-hungry machine learning (ML) workloads. Photonic computing, or computing using light, is a disruptive technology that can bring orders-of-magnitude performance and efficiency improvement to AI/ML with its ultra-fast speed, high parallelism, and low energy consumption. There has been growing interest in using nanophotonic processors for performing optical neural network (ONN) inference operations, which can make transformative impacts in future datacenters, automotive, smart sensing, and intelligent edge. However, the substantial potential in photonic computing also brings significant design challenges, which necessitates a cross-layer co-design stack where the circuit, architecture, and algorithm are designed and optimized in synergy.
In this talk, I will present my exploration to address the fundamental challenges faced by optical AI and to pioneer a hardware/software co-design methodology toward scalable, reliable, and adaptive photonic neural accelerator designs. First, I will delve into the critical area scalability issue of integrated photonic tensor units and present specialized photonic neural engine designs with domain-specific customization that significantly “compresses” the circuit footprint while realizing comparable inference accuracy. Next, I will present efficient on-chip training frameworks to show how to build a self-learnable photonic accelerator and overcome the robustness and adaptability bottlenecks by directly training the photonic circuits in situ. Then, I will introduce how to close the virtuous cycle between photonics and AI by applying AI/ML to photonic device simulation. Our proposed neural operator-based PDE solving framework enables real-time optical simulation with a 200x speedup over numerical solvers. In the end, I will conclude the talk with future research directions of emerging domain-specific photonic AI hardware with an intelligent end-to-end co-design & automation stack and deploying it to support real-world applications.
Biography:
Jiaqi Gu is a final-year PhD candidate in the Department of Electrical and Computer Engineering at The University of Texas at Austin, advised by Prof. David Z. Pan and co-advised by Prof. Ray T. Chen. Prior to UT Austin, he received his B.Eng. from Fudan University, Shanghai, China, in 2018. His research interests include emerging post-Moore hardware design for efficient computing, hardware/software co-design, photonic machine learning, and AI/ML algorithms.
He has received the Best Paper Award at the ACM/IEEE Asian and South Pacific Design Automation Conference (ASP-DAC) in 2020, the Best Paper Finalist at the ACM/IEEE Design Automation Conference (DAC) in 2020, the Best Poster Award at the NSF Workshop for Machine Learning Hardware Breakthroughs Towards Green AI and Ubiquitous On-Device Intelligence in 2020, the Best Paper Award at the IEEE Transaction on Computer-Aided Design of Integrated Circuits and Systems (TCAD) in 2021, the ACM Student Research Competition Grand Finals First Place in 2021, and Winner of the Robert S. Hilbert Memorial Optical Design Competition in 2022.
Towards Utility-Driven Data Analytics with Differential Privacy
Han Wang,
Illinois Institute of Technology
When:
11 a.m.–12 p.m., Thursday, February 23
Where:
435 Min H. Kao Building
Abstract:
With the increasing amounts of personal data being collected by different organizations for diverse set of applications, many privacy models have been proposed to sanitize the collected data for analysis without compromising individual privacy. In this talk, I will present our novel differentially private techniques that ensure provable privacy guarantees for the analyses on high dimensional and unstructured data such as textual, geo-spatial, and video data.
Specifically, differential privacy has been widely deployed to protect the privacy of individuals against attackers with arbitrary background knowledge. Our proposed differentially private techniques make the data or analysis results indistinguishable against adversarial data recipients or untrusted servers via minimum randomized data perturbation. Compared to other works, our approaches significantly improve the utility in different applications (e.g., topic mining on textual data, location-based services, and video analysis) while guaranteeing the same degree of privacy protection. Finally, I will present some of our ongoing and future works on fundamentally improving the privacy-accuracy tradeoff for privacy preserving machine learning.
Biography:
Han Wang is currently a PhD candidate in the Department of Computer Science at Illinois Institute of Technology, advised by Professor Yuan Hong. She got her MS degree in Software Engineering from Huazhong University of Science and Technology, China. Her research focuses on security and privacy (e.g., differential privacy, local differential privacy, secure multiparty computation, and applied cryptography), as well as trustworthy machine learning (e.g., robustness, privacy, and fairness). She has published numerous research results in top security and data science conferences and journals, such as S&P/Oakland, CCS, KDD, PETS, ICDE, EDBT, ICDCS, TDSC and TKDE.
Towards Automatic and Principled Vulnerability Discovery in the Linux Kernel
Hang Zhang,
Georgia Institute of Technology
When:
11 a.m.–12 p.m., Friday, February 24
Where:
435 Min H. Kao Building
Abstract:
The operating system kernel security is critical for the overall system reliability since the kernel runs in the highly privileged mode and is often a part of the trusted computing base. The kernel vulnerabilities can cause severe consequences if exploited by attackers (e.g., from denial-of-service to the whole system take-over). Thus, it is crucial to discover the vulnerabilities timely. Nowadays, many real-world vulnerabilities are discovered by manual code audits or dynamic testing (e.g., fuzzing). However, these approaches are either unscalable or of a random exploration nature. In this talk, I will discuss our research towards automatic and principled kernel vulnerability discovery with static program analysis. More specifically, I will first introduce how we can precisely identify known (but unpatched) security vulnerabilities in the downstream kernel binaries with FIBER – a tool we developed to capture and match the essential syntactic and semantic features of a security patch. Second, I will describe how we can statically discover specific unknown vulnerabilities of complex control and data flows, namely high-order taint-style vulnerabilities, in the kernel. This is enabled by our tool SUTURE, capable of precise and efficient high-order taint flow analysis. In the end, I will also briefly discuss some of my future research plans.
Biography:
Hang Zhang is a postdoctoral researcher at Georgia Tech. Previously, he obtained his PhD degree in computer science from the University of California, Riverside. His research area is system and software security, with an interest in discovering and analyzing security vulnerabilities in complex system software (e.g., the operating system kernels), especially with automatic program analysis techniques. His research has led to publications in top-tier computer security conferences (e.g., CCS, Usenix Security, S&P) and software engineering conferences (e.g., ICSE, FSE), multiple kernel vulnerabilities discovered, fixed, acknowledged, and sometimes rewarded, and open-source tools attracting interests from both academia and industry.
Marrying Application-Level Opportunities with Algorithm-Hardware Co-Design Towards Ubiquitous Edge Intelligence
Yang Zhao,
Georgia Institute of Technology
When:
11 a.m.–12 p.m., Monday, February 27
Where:
435 Min H. Kao Building
Abstract:
The recording-breaking performance of artificial intelligence (AI) algorithms, especially deep neural networks (DNNs), has motivated a growing demand for bringing powerful AI-powered intelligent functionalities onto edge devices, e.g., virtual reality/augmented reality (VR/AR) and medical devices, towards ubiquitous edge intelligence. However, the powerful performance of AI algorithms comes with much increased computational complexity and memory storage requirements, which stand at odds with the limited compute/storage resources on edge devices. Additionally, the stringent application-specific requirements, including real-time response (i.e., high throughput/low latency), high energy efficiency, and small form factor, further aggravate the aforementioned gap.
In this talk, I will introduce a holistic solution from energy- and latency-efficient architectures, to chips, and to integrated systems to closing the above-mentioned gap to enable more extensive AI-powered edge intelligence. Excitingly, my work shares the same underlying design insight, which is to advocate simultaneously harmonizing dedicated algorithms and hardware architectures via algorithm-hardware co-design while leveraging application-level opportunities to minimize redundancy within the processing pipeline and thus boost the achievable efficiency.
First, I will introduce our algorithm-hardware co-design work, called SmartExchange, which trades higher-cost memory storage/accesses for lower-cost computations to boost energy- and latency-efficiency. Motivated by the promising efficiency achieved by SmartExchange, we further validated its co-designed architecture by designing an AI acceleration chip prototype, which minimizes both the chip area and control overhead. To demonstrate the real-world advantages of the above SmartExchange architecture and its chip prototype, we developed i-FlatCam, a first-of-its-kind real-time eye-tracking system towards next-generation VR/AR devices, where we leverage the application-level opportunities to reduce both spatial and temporal redundancy. After that, we went beyond to build a scaled-up eye-tracking system, called EyeCoD, which targets a more general eye-tracking solution at the cost of marginally increased chip area as compared with i-FlatCams. Finally, I will conclude my talk with exciting future directions.
Biography:
Yang (Katie) Zhao is currently a postdoctoral fellow at Georgia Institute of Technology, and obtained her Ph.D. degree from Rice University in 2022. Her research expertise spans both computer architecture and domain-specific acceleration chip designs, with her research interest centering around enabling AI-powered intelligent functionalities on resource-constrained edge devices through a holistic solution from efficient architectures, to chips, and to integrated systems. Her research has led to over 20 publications (11 as the 1st/co-1st author) in top-tier computer architecture and circuit design conferences/journals, including ISCA, VLSI, MICRO, HPCA, ICCAD, ICASSP, FPGA, TNNL, TCAD, TVLSI, etc. Her work has been selected for IEEE Micro’s Top Picks of 2023 and won 1st place demonstration at the 32nd ACM SIGDA University Demonstration at DAC 2022, and she is a recipient of the 2020 Cadence Women in Technology Scholarship.
Student-Centered Machine Learning: Is Self-Supervision the Path toward General Computer Vision?
Henry Medeiros,
University of Florida
When:
11 a.m.–12 p.m., Thursday, March 2
Where:
435 Min H. Kao Building
Abstract:
Self-supervised learning techniques are rapidly becoming a keystone resource in the design of high capacity artificial intelligence models in various domains. However, the potential of such techniques to address many longstanding computer vision problems remains largely untapped. This talk discusses the application of self-supervised learning to the problem of image segmentation in two very distinct application domains. First, we present a panoptic segmentation algorithm for multiple species of fruit tree flowers for agricultural automation applications. We demonstrate that incorporating local appearance models and data augmentation strategies into a Monte Carlo framework makes it possible generate increasingly more accurate pseudo-labels for self-supervised model training. Our method accurately segments different species of flowers, most of which have significantly different appearances, substantially outperforming state-of-the-art models despite not being provided a single labeled picture of the flower species under consideration. We then show how such self-supervision techniques can be used to track passengers and baggage items in an airport security application. We show how data augmentation techniques can be incorporated into the architecture of modern instance segmentation methods to leverage semantic information produced by the model during pseudo-label generation. Specifically, modules commonly used to perform local aggregation in segmentation models, such as region of interest heads, offer a natural mechanism to integrate information generated by multiple inferences on augmented input data. We tested our model on a dataset comprising videos from multiple overhead security cameras. Unlike natural image datasets, the overhead perspective introduces dramatic appearance changes in the passengers, especially as they interact with their belongings. Our model is able to handle these challenges, performing on par with models trained in a fully supervised manner. These case studies point to promising future directions for the development of self-teaching robotic perception systems that would overcome many of the technical challenges preventing the application of robotic automation tools to unconstrained environments.
Biography:
Henry Medeiros is an Associate Professor of Agricultural and Biological Engineering at the University of Florida. His current research focuses on machine learning and stochastic computer vision models and their application to robotic perception systems. He has co-authored over fifty peer-reviewed publications and holds several U.S. and international patents. Before joining the University of Florida, he was an Associate Professor of Electrical and Computer Engineering at Marquette University and previously the Chief Technology Officer of Spensa Technologies, an agricultural automation technology start-up. He is a senior member of the IEEE and a recipient of the National Science Foundation Faculty Early Career Development Program award. He received his Ph.D. from the School of Electrical and Computer Engineering at Purdue University as a Fulbright scholar.
Homotopy and the foundations of formal verification
Carlo Angiuli,
Carnegie Mellon University
When:
11 a.m.–12 p.m., Friday, March 3
Where:
435 Min H. Kao Building
Abstract:
Dependent type theory is an expressive type system that lies at the heart of many interactive theorem provers, tools for writing machine-checked proofs of program correctness. In recent years, researchers in the emerging field of homotopy type theory have uncovered surprising connections between dependent type theory and a branch of mathematics known as homotopy theory.
In this talk, I will show how theoretical ideas from homotopy type theory address shortcomings in the practice of formal verification. First, I will discuss how homotopy type theory lets us deduce representation independence theorems stating that two implementations of an abstract interface can be exchanged freely, provided they have equivalent input/output behavior. These theorems do not hold in traditional theorem provers, but can be leveraged to provide modular techniques for verifying data structures. Then, I will present Cartesian cubical type theory, a dependent type theory which reconstructs key features of homotopy type theory in the type system of a functional programming language. My collaborators and I have developed several prototype implementations of Cartesian cubical type theory, and a closely-related theory has been implemented at scale in the Agda programming language; experiments indicate that these systems are significantly easier to use than earlier implementations of homotopy type theory.
Biography:
Carlo Angiuli is a postdoctoral fellow at Carnegie Mellon University, where he obtained his PhD in 2019 under the supervision of Robert Harper. His research interests are in programming languages and formal verification, with an emphasis on dependent type theory and its connections with mathematics and logic. His work has won the School of Computer Science Distinguished Dissertation Award at CMU and the Best Paper Award by Junior Researchers at FSCD.
Enabling Visualization of Data Analysis Programs
Rebecca Faust,
Virginia Tech
When:
11 a.m.–12 p.m., Monday, March 6
Where:
435 Min H. Kao Building
Abstract:
Data analysis methods inherently lose information when moving from initial data to final results, such as context from the underlying data features. The loss of such information inhibits analysts’ ability to reason about the processes performed on their data and interpret the results. In this talk, I will discuss my efforts to design visualization and interaction methods that re-expose some of the information lost during analysis. First, I will focus on using visualization to re-introduce underlying data features in dimension reductions. I will demonstrate how my DimReader method provides lost context of the underlying data features via gradient visualization. Additionally, I will discuss how explainable, interactive projections provide context while enabling data exploration through feature-based reorganization. Then, I will broaden the scope and illustrate how visualizing analysis scripts, via my Anteater method, provides context from the internal script data that illuminates how the analysis reaches its final results.
Biography:
Rebecca Faust is a postdoctoral researcher and Computing Innovations Fellow in the Sanghani Center for AI and Data Analytics at Virginia Tech working with Chris North. She received her PhD from the University of Arizona, where she was advised by Carlos Scheidegger. Her research interests lie at the intersection of data visualization and data analysis, with an emphasis on designing interactive visualization methods to enable explainability and interpretability of analysis methods and results.
Wearable Technology for Healthcare and Athletic Performance
Amanda Watson,
University of Pennsylvania
When:
11 a.m.–12 p.m., Tuesday, March 7
Where:
435 Min H. Kao Building
Abstract:
In the future, wearable technology will provide a continuous, autonomous, and comprehensive assessment of a person’s health. This future presents a tremendous opportunity for the development of new wearable devices that will support predictive analytics and personalized medicine. On the hardware side, we are liberating sensors from the bench top in the clinical lab and introducing them into the wild. Most commonly, smartwatches provide sensors that enable vital sign monitoring, such as heart rate and pulse oximetry. However, many of the most powerful sensing techniques are still limited to the benchtop. In this talk, I will introduce my recent work on developing a wearable optical spectroscopy device for non-invasive health monitoring. I will discuss the challenges of redesigning traditional spectroscopic methods to overcome the resource constraints of a wearable device and facilitate the adaptation to dynamically changing environments. Further, I will explore how this device can be used to continuously monitor biomarkers such as glucose or opioids. Finally, I will discuss my vision for the future; enabling a future where everyone has a comprehensive understanding of their own health at their fingertips.
Biography:
Amanda Watson is a Postdoctoral Scholar in the Department of Computer and Information Science at the University of Pennsylvania.
Watson is a member of the PRECISE center mentored by Insup Lee and James Weimer. Her research seeks to develop new wearable sensors that provide high-quality longitudinal datasets, the underlying infrastructure to support these devices, and its applications in healthcare and athletic performance. Her work has appeared at top tier venues including ACM IMWUT and IPSN and she was selected as a Rising Star in EECS by MIT, and a Rising Star in CPS by the University of Virginia. Previously, she was a member of the LENS lab at William & Mary, where she researched wearable technology with Gang Zhou.
Watson’s research interests include wearable technology, medical cyber-physical systems, internet of medical things, ubiquitous and mobile computing, and smart health.
Towards Next-Gen Cyber-Physical Infrastructure for Urban Environments
Weizi Li,
University of Memphis
When:
11 a.m.–12 p.m., Wednesday, March 8
Where:
435 Min H. Kao Building
Abstract:
The world is rapidly urbanizing, with two-thirds of the global population projected to be living in cities by 2050. This further strengthens the role of cities as hubs for transportation, information, and socio-economic development. However, this growth also brings many challenges such as traffic congestion, pollution, and economic disparity. While cities worldwide have developed smart city initiatives, exact paths to solving existing urban issues remain elusive. In this talk, I will introduce our efforts in developing next-gen, cyber-physical infrastructure for cities from the Intelligent Transportation System and Network Science perspectives. These infrastructures carry the potential to advance efficiency, intelligence, and resilience of our cities. I will present three projects. First, I will discuss our recent advances in controlling hybrid traffic (a mixture of robot and human-driven vehicles) at complex and unsignalized intersections [1]. Second, I will describe our efforts to reconstruct citywide traffic dynamics via mobile data [2], which pave the road for citywide hybrid traffic control. Finally, I will introduce a new learning-based framework for graph sparsification [3], which is expected to be versatile as increasingly complex networks are formed in cities due to ubiquitous sensors and hyperconnectivity.
- Learning to Control and Coordinate Hybrid Traffic Through Robot Vehicles at Complex and Unsignalized Intersections
- City-Scale Traffic Animation Using Statistical Learning and Metamodel-based Optimization
- A Generic Graph Sparsification Framework using Deep Reinforcement Learning
Biography:
Weizi Li is Assistant Professor of Computer Science at the University of Memphis. He was a Michael Hammer Postdoctoral Fellow at MIT and he received his PhD in computer science from the University of North Carolina at Chapel Hill. He has also spent time at the University of Maryland, Disney Research, Lenovo Research, National University of Singapore, and George Mason University.
His research interests include intelligent transportation systems, robotics, reinforcement learning, and multi-agent simulation. He is passionate about using simulation and learning to design and shape the cities of the future. He is a recipient of the NSF CRII Award.
Towards a Hardware/Software Co-Design Approach for CPS and IoT Security
Anomadarshi Barua,
University of California, Irvine
When:
11 a.m.–12 p.m., Thursday, March 9
Where:
435 Min H. Kao Building
Abstract:
Have you ever wondered how simple music can leak deadly pathogens from a bio-safety lab? Have you ever considered how an industrial robotic arm can be compromised from a cloud infrastructure? Have you ever imagined an attacker could use a magnetic field to shut down a weak microgrid? If the above questions excite you, please attend my research seminar, where my research answers the above questions with a practical demonstration of CPS attacks on industry-used infrastructure. With this seminar, the audience will gain a deep insight into how hardware/software interaction in CPSs can open a “Pandora’s Box” of unknown threats that can come from very unconventional ways.
My research sits at the intersection of hardware and software layers in CPSs and IoTs, exploring how interactions of cyber and physical components can affect the safety and controllability of closed-loop control from sensing to actuation. Most attacks on CPSs and IoTs can be propagated from the physical domain to the cyber domain or vice-versa and hence, can be termed as cross-domain attacks. To understand these cross-domain attacks, a very different set of methodologies and tools is needed. My research talks about how these tools can be made. Moreover, the seminar will show how cross-domain vulnerabilities can be handled using new hardware/software co-design approaches to detect, contain and isolate threats in CPSs and IoTs.
Biography:
Anomadarshi Barua is a PhD candidate studying computer engineering at the University of California Irvine. He received his MS degree in embedded computing systems jointly at the University of Southampton (UoS) and the Norwegian University of Science and Technology (NTNU) in 2016. Now, he is working under the supervision of Professor Al Faruque in the Autonomous and Intelligent Cyber-Physical Systems (AICPS) Lab. His current research focuses on the system-level design of embedded and Cyber-Physical-Systems (CPSs) with a special interest in CPS security, low-power design, and computer microarchitecture. He publishes in the security conferences like ACM CCS, USENIX, ACSAC, RAID, and CHES and system conferences, like ICCPS and ICCAD. He also has worked in several industries in North America and Europe, such as Intel Corporation (USA), Solidigm (USA), Nordic Semiconductor (Norway), and IDEAS (Norway). In his leisure time, he likes to travel to new places and go for hiking.
Automatically Preventing, Detecting, and Repairing Crucial Errors in Programs
Jialu Zhang,
Yale University
When:
11 a.m.–12 p.m., Friday, March 10
Where:
435 Min H. Kao Building
Abstract:
High-impact errors in programs cause huge money losses, are notoriously expensive for programmers to repair, and affect millions of real-world users. In this talk, I will highlight the need to move beyond our post-mortem, manual error handling to develop tools to automatically prevent, detect, and repair program errors. I will present my methods and findings on misconfiguration detection, merge conflict resolution, and feedback generation for errors in students’ programming assignments. The tools I have developed combine a variety of new techniques including program analysis, machine learning, and large language models. I will present my progress, and discuss the implications for the future of error detection and repair research.
Biography:
Jialu Zhang is a final-year PhD student in computer science at Yale University advised by Ruzica Piskac. He develops tools to automatically prevent, detect, and repair high-impact errors in systems (misconfigurations), collaborative software development (merge conflicts and continuous integration errors), and most recently, CS Education (generating feedback for intro-level and competitive-level programming assignments). His tool development delivers practical impact. For example, the detected misconfigurations have been confirmed and resolved by the original developers on Github, and his merge conflict resolution tool is under productization in the daily used web browser: Microsoft Edge. Previously, he spent two summers at Microsoft Research (MSR) working with Shuvendu Lahiri, Sumit Gulwani and Jose Cambronero. He received his BS degree from Shanghai Jiao Tong University (IEEE Honor Class).
Video recordings of the faculty candidate seminars are available on the EECS Portal site, in the folder “EECS Faculty Candidate Recordings- Feb-Mar 2023,” located inside the Documents folder.