Past SIAI Seminar Series

Ayoung Kim, Seoul National University

Integrating Learning-Based Methods with Model-Based Approaches in Spatial Intelligence

APRIL 13, 2023

Abstract

This talk will cover three recent studies from our group that explore integrating learning-based methods into existing model-based approaches in spatial intelligence, also known as Simultaneous Localization and Mapping (SLAM). Recognizing a place is crucial, and recent advancements in learning-based methods have demonstrated superior performance. First, the study revisits the approach of modeling images as documents and comprehending images through visual words. By considering a line in an image as a sentence in a document, place recognition can be viewed as a sentence-to-sentence matching problem. Secondly, the study evaluates the manipulation of SLAM as a problem by combining an AI-based module and graph SLAM. Doing so necessitates predicting not only the mean value but also the uncertainty for the predicted value. Lastly, the study investigates other sensor modalities beyond the visible spectrum. One major challenge for these sensors is the scarcity of well-labeled data. The study introduces an image-to-image translation and presents a technique for labeling thermal images.

Biography

Ayoung Kim works as an associate professor in the department of mechanical engineering at Seoul National University since Sep. 2021. Before joining SNU, she was at the civil and environmental engineering, Korea Advanced Institute of Science and Technology (KAIST) from 2014 to 2021. She has B.S. and M.S. degrees in mechanical engineering from SNU in 2005 and 2007, and an M.S. degree in electrical engineering and a Ph.D. degree in mechanical engineering from the University of Michigan (UM), Ann Arbor, in 2011 and 2012. She also worked as a post-doctoral researcher in naval architecture and marine engineering at UM in 2013 before she worked at Electronics and Telecommunications Research Institute (ETRI) as a senior researcher.

Sanjeev Koppal, Ph.D. University of Florida

Adaptive Attention: Bringing Active Vision into the Camera

NOVEMBER 17, 2021

Abstract

Most cameras today capture images without considering scene content. In contrast, animal eyes have fast mechanical movements that control how the scene is imaged in detail by the fovea, where visual acuity is highest. The prevalence of active vision during biological imaging, and the wide variety of it, makes it very clear that this is an effective visual design strategy. In this talk, I cover our recent work on creating *both* new camera designs and novel vision algorithms to enable adaptive and selective active vision and imaging inside cameras and sensors.

Biography

Sanjeev Koppal is an Associate Professor at the University of Florida’s Electrical and Computer Engineering Department. He also holds a UF Term Professor Award for 2021-24. Sanjeev is the Director of the FOCUS Lab at UF. Prior to joining UF, he was a researcher at the Texas Instruments Imaging R&D lab. Sanjeev obtained his Masters and Ph.D. degrees from the Robotics Institute at Carnegie Mellon University. After CMU, he was a postdoctoral research associate in the School of Engineering and Applied Sciences at Harvard University. He received his B.S. degree from the University of Southern California in 2003 as a Trustee Scholar. He is a co-author on best student paper awards for ECCV 2016 and NEMS 2018, and work from his FOCUS lab was a CVPR 2019 best-paper finalist. Sanjeev won an NSF CAREER award in 2020 and is an IEEE Senior Member. His interests span computer vision, computational photography and optics, novel cameras and sensors, 3D reconstruction, physics-based vision, and active illumination.

Lakshmi Theagarajan, Indian Institute of Technology Palakkad

Index Codes for Cyber Physical Systems: Machine Learning Meets Source Coding

DECEMBER 9, 2019

Abstract

We shall discuss the problem of constructing index codes that maximize a data network's throughput and minimize power consumption. In this setup, multiple users, with some side information, demand certain subsets of data from a central node. This scenario is the basic model of the data sharing network in various cyber-physical systems and connected networks. The goal of the index code is to minimize the required transmission rate while ensuring that the users can recover the demanded data using an encoded broadcast message from the server. Thus, minimizing the power expended for communication while maximizing throughput. We shall discuss a generalized version of the index coding problem, where both the side information and user-demanded data can be coded. The problem of index code construction, for a given set of side information and users’ demands, can be modeled as a matrix completion problem. Traditional machine learning algorithms used for matrix completion do not take advantage of the inherent structure in the index codes to construct efficient index codes. We shall discuss novel techniques proposed by us for low-rank factorization of structured matrices. We shall see that the proposed machine learning methods for structured matrix completion construct efficient index codes compared to the traditional methods known in the literature so far.

Biography

Dr. Lakshmi N. Theagarajan obtained his MS and PhD from the Indian Institute of Science in 2015. In his doctoral research, he developed novel low-complexity probabilistic graphical model based inference algorithms that made massive MIMO signal processing methods practically realizable. In 2016-2017, he was a postdoctoral researcher at the Sensor Fusion Lab at Syracuse University, where he developed novel online statistical learning algorithms for distributed detection and estimation problems. Currently, he is an assistant professor at the Indian Institute of Technology Palakkad and a visiting assistant professor at McMaster University. He has also worked in Cisco Systems R&D and National Instruments R&D in the past. His areas of research include statistical learning and inference in large-scale wireless networks, distributed machine learning (federated learning), sparse signal processing, design of optimal modulation and coding schemes, visible light communication, information and coding theory.

Martin Slawski, George Mason University

Learning from Integrated Data under Mismatch Corruption

NOVEMBER 15, 2019

Abstract

Contemporary data acquisition and analysis frequently involves the integration of multiple pieces of information about a common set of entities into a single comprehensive data set. In the absence of unique identifiers, merging corresponding fragments of data can be demanding and error-prone. In this talk, several techniques accounting for mismatched data in regression setups with features in one file and labels in another file will be presented. The problem is formulated in terms of an unknown permutation of the labels. While recovering the permutation tends to be challenging from computational and statistical viewpoints, learning the regression relationship is often feasible under additional prior knowledge that is commonly available in practice. Connections to data privacy and linkage attacks will be briefly discussed as well.

Biography

Martin Slawski is an assistant professor in the Department of Statistics at George Mason University. This fall semester, he is a visiting faculty at Columbia University. Before his current appointment, he spent two years as a postdoc at Rutgers University after graduating with a PhD in Computer Science from Saarland University, Germany. His main research interests include structured and compressed representations of high-dimensional data, record linkage and data integration, and optimization in statistical settings. His research is funded by NSF and the National Institute of Justice.

Ceren Budak, University of Michigan, Ann Arbor

Examining Political Information and Behavior Online

OCTOBER 14, 2019

Abstract

With an increasing amount of data available online, we are now able to examine political information and behaviors through a new lens. In this talk, I will cover a series of studies that underline this promise for the study of news producers, citizens, and social movement organizations. First, focusing on the news media, I will characterize the spread of fake news during the 2016 Presidential elections. Through the use of heterogenous data, I will examine the interplay between news media production and consumption, social media behavior, and the information the electorate retained about the presidential candidates leading up to the election. Second, turning to the citizens, I will examine how individuals conform to community norms in political discussions. Past research identifies many processes that contribute to maintaining stable norms, including self-selection, pre-entry learning, post-entry learning, and retention. What is the relative importance of these processes? I will answer this question through an analysis of political subreddits on Reddit with stable and distinctive toxicity levels. Finally, by building predictive models to detect social movement organizations (SMOs) at scale, I will examine SMO participation in online social and political movements.

Biography

Ceren Budak is an Assistant Professor of Information and Assistant Professor of Electrical Engineering and Computer Science, College of Engineering at the University of Michigan, Ann Arbor. Her research interests lie in the area of computational social science. She utilizes network science, machine learning, and crowdsourcing methods and draws from scientific knowledge across multiple social science communities to contribute computational methods to the field of political communication.

Anastassia Loukina, Educational Testing Services

Algorithmic Fairness and Bias in Automated Test Scoring

SEPTEMBER 19, 2019

Abstract

The issues of algorithmic fairness and bias have recently featured prominently in many publications highlighting the fact that training the algorithms for maximum performance may often result in predictions that are biased against various groups. Educational applications based on NLP and speech processing technologies often combine multiple complex machine learning algorithms and are thus vulnerable to the same sources of bias as other AI systems. Yet such systems when used in assessment can have high impact on people's lives. In this talk I will discuss the issues of fairness and bias in educational applications using as an example the case of automated scoring of non-native English spontaneous speech.

Biography

Anastassia Loukina is a research scientist in the Research and Development division at Educational Testing Service (ETS) in Princeton NJ. ETS develops, administers and scores more than 50 million tests annually in more than 180 countries at more than 9,000 locations worldwide. The NLP & Speech group at ETS develops technology for automated scoring of open ended items, classroom support tools that teachers, and tools that can aid in the test development process. Since joining ETS, Anastassia has led the research to improve the validity, reliability and fairness of speech-based educational application and made key contribution to the successful launches of several automated scoring systems. She published more than 40 papers and book chapters, holds several patents and frequently attends international conferences and workshops.

Stephanie Dinkins, Artist

Artificial Intelligence, Race, Gender & Future Histories

APRIL 9, 2019

Biography

Stephanie Dinkins is a transdisciplinary artist who creates platforms for dialog about artificial intelligence as it intersects race, gender, and our future histories. Her art employs lens-based practices, the manipulation of space, and technology to grapple with notions of consciousness, agency, perception, and social equity. She is particularly driven to work with communities of color to develop AI literacy and co-create more inclusive, equitable artificial intelligence. Dinkins’ artwork is exhibited internationally at a broad spectrum of community, private and institutional venues – by design. These include International Center of Photography, NY, Bitforms Gallery, NY, Miller Gallery, Carnegie Mellon University, Institute of Contemporary Art Dunaujvaros, Herning Kunstmuseum, Spelman College Museum of Fine Art, Contemporary Art Museum Houston, Wave Hill, the ‘Studio Museum in Harlem, Spedition Bremen, and the corner of Putnam and Malcolm X Boulevard in Bedford-Stuyvesant, Brooklyn.

Anup Kalia, Jin Xiao and Maja Vukovic, IBM Research

AI-Infused Service Management and Optimization

FEBRUARY 28, 2019

Abstract

Trends in big data and emergence of cognitive computing are powering intelligence in cloud platforms. Cognitive cloud platforms emerge as systems which embody human-like reasoning in order to accelerate development and dynamic adaptation of novel applications and cloud-enabled processes.In this talk, we present an overview of AI innovations in IT service management, ranging from problem determination to compliance. We then focus on service request management, and demonstrate how the end-user and support teams can interact with the system through natural language interfaces, to resolve problems and/or request service changes and adapt policy configurations.

Biography

Anup Kalia is a research staff member of the Cognitive Service Management organization at IBM T. J. Watson Research Center, NY. His research interests include service computing, multiagent systems, cognitive science and software engineering. In IBM Research, he is exploring different techniques such as text mining, natural language processing, machine learning and multiagent-based business process modeling to solve problems in the areas of service analytics, automation, knowledge graph extraction, DevOps and Blockchain. Kalia is a member of IEEE and ACM.Jin Xiao is a research staff member at IBM T. J. Watson Research Center, where he is a member of the cognitive service management research department. His research interests include service automation, knowledge generation, NL-to-API interactions, data mining in network and system logs for analytics, and more recently on change management in DevOps and micro-services. Xiao is a member of IEEE. Contact him at [email protected] Vukovic is a research manager and a research staff member at IBM T. J. Watson Research Center, where she leads the cognitive service management research department. Her research interests include services computing, software engineering, IT service management and AI planning. Maja is a senior member of IEEE, member of IBM Academy Technology, and an IBM Master Inventor. Contact her at [email protected].

Ching-Hua Chen, IBM Research

Computational Health Behavior Research at IBM Research

DECEMBER 7, 2018

Abstract

Dr. Ching-Hua Chen will describe ongoing health behavior research at IBM Research. Examples of research topics include physical activity coaching, continuous stress monitoring and care management. On the topic of physical activity coaching, she will share how a combination of qualitative studies and machine learning inspired a concept for an adaptive intervention for promoting physical activity. On the topic of continuous stress monitoring, she will share how wearable sensing technologies were used to understand how context and physiology may allow us to infer a user's perceived stress level. Finally, on the topic of care management, she will share how behavioral phenotyping and causal inference methods can be used to help care managers better manage their limited time.

Biography

Dr. Chen is a manager and research scientist at IBM Research in Yorktown Heights, NY. Her team develops and tests technology-enabled approaches to understanding health behavior and patient decision-making. She is investigating methods for increasing the speed and accuracy with which practitioners can personalize interventions for health behavior modification. To this end, she is interested in quantifying health behavior and patient experience through the analysis of various types of temporal data (e.g., medical/health records, social platforms, wearable devices and IoT sensors). She has a dual-title Ph.D. in business administration and operations research from Penn State University.

Philip Gautier, Amazon

Infinitely Scalable Optimization for Supervised Learning: Practical Approaches at Amazon

NOVEMBER 29, 2018 

Abstract

Industrial applications of machine learning are increasingly making use of commodity hardware to process ever-larger datasets. In this talk, I'll address the question: How do we design optimal software systems for efficiently solving general supervised learning problems when the data are arbitrarily large in terms of dimension and number of examples? I will advocate stochastic gradient descent and a distributed parameter server as solutions, with theoretical and practical justification. I will also discuss approaches to hyperparameter optimization and statistical inference at "internet scale."

Biography

Philip Gautier is a machine learning researcher and practitioner with interests in deep learning, statistical modeling, convex and non-convex optimization, systems and software design for scalable machine learning, applied data analysis and data visualization. He has experience in academia and industry applying machine learning to supply chain optimization, forecasting, scalable system design and computer vision. He received his Ph.D. in statistics at Purdue University under William S. Cleveland, with research in the field of distributed convex optimization for parameter inference on maximum likelihood models under the MapReduce paradigm.