Computer Science Department

CS Seminar Series

Spring 2021

  • 4/8 Esra Ataer-Cansızoğlu (1:20pm) -- Deep Ranking for Style-Aware Product Recommendation in Interior Design

    Deep Ranking for Style-Aware Product Recommendation in Interior Design
    Esra Ataer-Cansızoğlu
    Facebook

    When: Thu, 4/8, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    We present a deep learning based room image retrieval framework that is based on style understanding. Given a dataset of room images labeled by interior design experts, we map the noisy style labels to comparison labels. Our framework learns the style spectrum of each image from the generated comparisons and makes significantly more accurate recommendations compared to discrete classification baselines.

    Bio:
    Esra Cansizoglu is a machine learning engineer at Facebook working on geospatial image analysis for mapping. She received her PhD in Electrical Engineering from Northeastern University and her MS in Computer Science from Boston University. Her research interests are in Computer Vision and Machine Learning. She co-authored several patents and publications in peer-reviewed conferences and journals. Her experience in research and development spans various domains including medical image processing, robotics and recommendation systems.

  • 4/15 Huseyin Demirci (1:20pm) -- The Next Generation Sequencing Era: Applications, Challenges and Threats

    The Next Generation Sequencing Era: Applications, Challenges and Threats
    Huseyin Demirci
    University of Luxembourg

    When: Thu, 4/15, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    In this talk we will briefly introduce the genomics sequencing era. With the availability of genome sequencing, new applications for diagnosis and therapy are now possible. We will give examples from health, biotechnology and food industries as applications of next generation sequencing. Genome sequencing is inevitably becoming a part of our lives. We will present our results from a 1.100 exome project, which helped to identify many disease-causing mutations for rare diseases and diagnose patients. We will point out software tools assisting with this type of research. We will finally mention the privacy issue of the genomics data which is becoming an increasingly important topic.

    Bio:
    Huseyin Demirci received his PhD degree from Marmara University, Turkey in 2004 with the thesis "Stochastic Analysis of Block Ciphers". From 1998 to 2016 he worked as a senior researcher at the National Scientific and Technological Research Center of Turkey. He worked on cryptology and cryptanalysis of symmetric ciphers including well known ciphers AES and RC4. He participated in national and EU projects on wireless and lightweight security. Between 2013-2016, he contributed to the foundation of first Genome Sequencing and Bioinformatics analysis Center of Turkey. Between 2016-2019 he worked as a post-doctoral researcher at the University of Minho, Portugal. Since 2020, he is a research associate at the University of Luxembourg. His research interests include genome sequence analysis, information security and genomics privacy.

  • 4/22 Can Yilmaz (1:20pm) -- How to be(come) a software engineer or data scientist

    How to be(come) a software engineer or data scientist
    Can Yilmaz
    Amazon

    When: Thu, 4/22, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Computer Science is a major where you learn the fundamentals of computing. Students sometimes can have hard times figuring out how to translate what they learned into skills that are desired in the current job market or what skills are the most important for their dream job. Computer science degrees can translate into many different job titles: software developer, data scientist, technical program/product manager etc. Although one can think almost all these titles are similar and about coding, they are quite different and coding is not a fundamental skill for some. In this talk, I will discuss software engineer and data scientist titles: what companies pay them to do, fundamental skills, salary expectations and most importantly, strategies to use in the job market to become one.

    Bio:
    Can is Software Development Engineer working at Amazon. He worked at Microsoft as a Data Scientist before. At Microsoft, Can helped bring Cortana to Outlook as a virtual assistant to help users with their daily email tasks. He also involved in the efforts to have Cortana and Alexa to talk to each other to give users a complete voice assistant experience where they can control their smart home and computers at once. At Amazon he is a part of a small team that creates a confidential platform that will be a breakthrough for AI oriented data management tools. Can is self-taught engineer and data scientist. He has a master degree in Computer Science and bachelors in Mathematics.

  • 4/29 Patrick McClure (1:20) -- A Probabilistic Perspective on Deep Neural Networks

    A Probabilistic Perspective on Deep Neural Networks
    Patrick McClure
    National Institute of Mental Health

    When: Thu, 4/29, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Deep neural networks (DNNs) are machine learning models that are being applied to increasingly diverse areas. Many of these models are used to make decisions that have substantial impact, such as those related to disease diagnosis from medical images or autonomous driving. However, can their predictions be trusted? A key component of the trustworthiness of a prediction model is proper estimation of prediction uncertainty. Improving DNN uncertainty estimation also leads to better performance for a variety of problems, including continual learning, distributed learning, and anomaly detection. Bayesian probability tools offer a principled solution for improving DNN uncertainty estimation by learning distributions of DNNs (i.e. Bayesian DNNs). In this talk, we will briefly introduce Bayesian DNNs and demonstrate their practical usefulness using neuroimaging examples.

    Bio:
    Dr. Patrick McClure is a Machine Learning Project Lead at the National Institute of Mental Health (NIMH). He received his PhD in Computational Neuroscience from the University of Cambridge under the supervision of Dr. Nikolaus Kriegeskorte. Before attending Cambridge, he received his MS in Computer Science and his BS in Bioengineering from the University of Louisville. Patrick’s research has primarily focused on developing deep learning and probabilistic modeling tools and applying them to the areas of computational neuroscience, medical image analysis, and computer vision.

  • 5/6 Iris Howley (1:20pm) -- Explaining AI for Decision-Making

    Explaining AI for Decision-Making
    Iris Howley
    Williams College

    When: Thu, 5/6, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Users of artificial intelligence (AI) decision-making systems rely on algorithms to help them make day-to-day decisions, but may not understand the potential flaws and biases of these algorithms, even if the algorithm is open and not blackbox. This talk provides an overview of explainable AI followed by a discussion of how our research group expands on current post-hoc methods for explainable AI. We use a method from the learning sciences and human-computer interaction communities, Cognitive Task Analysis (CTA), to identify what knowledge components comprise expert understanding of an algorithm. We apply CTA to Bayesian Knowledge Tracing, an AI algorithm commonly used in learning analytics systems, so that more systematic, rigorous post-hoc explanations for AI algorithms can be developed and evaluated.

    Bio:
    Iris Howley is a human-computer interaction, artificial intelligence, and learning science researcher focusing on enabling users to overcome obstacles to effective decision-making and community participation through the design of technologies. She received her B.S. in Computer Science from Drexel University and her M.S. and Ph.D. in Human-Computer Interaction from Carnegie Mellon University. Prior to becoming an Assistant Professor of Computer Science at Williams College in 2017, she was a postdoctoral research fellow with the LINK Research Lab at the University of Texas Arlington and the Lytics Lab at Stanford University. At the moment, she is researching the design and deployment of interactive explainables for users of algorithmic systems in educational contexts with the support of an NSF grant.

  • Tue 5/11 Poorna Talkad Sukumar (1:20pm) -- Towards a Realistic Understanding Of Personal Visualization

    Towards a Realistic Understanding Of Personal Visualization
    Poorna Talkad Sukumar
    Notre Dame

    When: Tue, 5/11, 1:20-2:20pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Personal data comprises any data that is relevant to one's personal life, such as, health and fitness data and social media interactions. Personal visualizations refer to interactive visual representations of personal data and present a separate class of visualizations with distinctive goals and usage characteristics. They aim to support self-reflection, are consumed in less formal contexts and often on mobile devices, and by people with different motivations, interests, and resources. However, studying personal visualizations can be challenging and traditional evaluation methods may not be suitable for assessing their distinctive goals. To study personal visualizations, it is important to study people interacting with their own data and in realistic settings.

    In this talk, I will present two user studies, employing different methods, conducted with a web-based interface presenting visualizations of the personal data gathered as part of a large-scale, longitudinal sensing study. The first study characterizes users' exploratory behaviors on the interface by analyzing logged interactions (mouse hovers and clicks) from 369 participants as they each explored their own data. The second study employs a think-aloud method to identify the personal insights gained by the participants, contextual information recalled by the participants to interpret their data, and usability issues with the interface. Coalescing the findings from the two studies, I will discuss possible design directions for improving personal visualizations and expanding their impact.

    Bio:
    Poorna Talkad Sukumar is a Ph.D. Candidate in the Department of Computer Science and Engineering at the University of Notre Dame. Her main areas of research are Human-Computer Interaction (HCI) and Information Visualization. She has worked on a number of diverse projects, including the mitigation of cognitive biases of admissions reviewers using visualization tools, exploration of research methods suitable for studying personal visualizations, improving equity in STEM classrooms using visualizations, and understanding and facilitating team behavior using unobtrusive sensing technologies. Her research focuses on emerging topics in HCI and visualization having broader impacts and she has published in top venues, including ACM CHI, EuroVis, and CSCW. She is generally interested in designing and building interactive visualization systems and well-versed in applying various human-centered methods to evaluate these systems.
    She also has a Master’s degree in Mobile and Ubiquitous Computing from Lancaster University, UK.

  • 5/13 Brian McInnis (1:20pm) -- Opportunities and barriers to facilitating evidence-based discussion about civic issues

    Opportunities and barriers to facilitating evidence-based discussion about civic issues
    Brian McInnis
    UC San Diego

    When: Thu, 5/13, 1:20-2:20pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    News websites have the potential to facilitate large scale evidence-based discussion about civic issues, like climate change, but the online discussions are often overrun by toxicity and misinformation. Even when people are not shouting at each other, online discussions about civic issues rarely focus on evidence, such as data and visualizations. In this talk, I will share our analysis of how people comment about climate change data in the discussion threads at three news websites (i.e., Breitbart News, the Guardian, the New York Times). The findings illustrate that data-centered talk, while rare, can provide valuable insights that could help center an online discussion around evidence—collection, analysis, and visual representation—to engage with an article's narrative. My research explores how techniques from crowdsourcing offer a potential way to promote and encourage evidence-based discussion. I will draw examples from my own work studying news websites, collaboration platforms, and in-the-wild civic initiatives to review critical system design opportunities and barriers.

    Bio:
    Dr. Brian McInnis is a postdoctoral scholar at UC San Diego and a member of the Design Lab. Brian explores how to help people collaboratively build insights around policy concerns. Brian earned his PhD in Information Science from Cornell University in 2019 where his thesis investigated how people build insights around policy concern through a series of studies that involved crowd workers in online discussions related to the AMT participation agreement. Prior to joining Cornell, Brian worked at the RAND Corporation, where he studied a range of public policy issues—from the design of youth summer learning programs to predictive policing techniques. Brian earned his Masters of Public Policy from Vanderbilt University's Peabody College of Education as well as a dual Bachelors in Economics and History from the University of California at Davis.

  • Mon 5/17 Uzma Mushtaque (1:50pm) -- Recommender System Models for Online Retail and Subscription Platforms

    Recommender System Models for Online Retail and Subscription Platforms
    Uzma Mushtaque
    RPI

    When: Thu, 5/17, 1:50-2:50pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Recommender Systems are used on all online platforms to provide personalized assortments to users. Well-known consideration-set theory from cognitive science identified information overload impacting user choice. In this research, a new family of discrete choice models called MNL-CE models is devised to establish a direct connection between size of the assortment offered and the probability of user selection or rejection. When used as inputs to assortment optimization problem, these models present an optimal assortment that should be offered to minimize no-choice (rejection) probability for a given user. The notion of incorporating diversity in assortments is also addressed by these models. These models find direct application in online retail (Amazon) and/or subscription platforms (Netflix) where customers are inundated with large assortments in the form of personalized recommendations. Algorithms devised to solve for optimal assortments are analyzed in-depth. Data collection and Machine Learning (ML) approaches to parameter estimation is an existing open research question which is discussed in detail. The models developed have their roots in random-utility theory, therefore there are three logical extensions of this research: 1) To devise explainable models for Recommender Systems in specific and for other applications in general, 2) Discrete choice models are naturally represented as Neural Networks, therefore an improved overarching technique for estimation can be developed using AI 3) Improve peer-to-peer systems by utilizing these models.

    Bio:
    Dr. Uzma Mushtaque is a Lecturer at the Department of Computer Science at Rensselaer Polytechnic Institute (RPI). She has a PhD in Decision Sciences and Engineering Systems from RPI along with a Postdoc experience on developing predictive choice models for various online applications. Her research interests include recommender systems, machine learning, AI, statistical learning, optimization, data analytics and operations research. Additionally, she has a master’s from Penn State's Business School and 3 years of professional experience in the information technology (IT) field working as a consultant in Oracle ERP implementation projects. One of her most recent industry experiences includes working as a Senior Data Scientist at one of the top consulting firms on credit and risk analytics. Her PhD thesis and published work includes an in-depth, iterative, and methodical exploration of some of the Big-Datasets available for consumer preference ratings. Her future research is aimed at estimating choice models using machine learning approaches by collecting relevant data. Currently she is also working on creating XAI (Explainable AI) models for recommender systems and models for personalization in education. Dr. Mushtaque's background in Engineering and Computer Science, work experience in Data Science consulting, a master's in supply chain management and her current research motivates the inter-disciplinary nature of her teaching and research interests.

Winter 2021

  • 1/21 Career Panel (5pm)

    Career Panel
    Featuring Laura Marlin '19, Akshay Kashyap '18, Julian Jocque '15, Eric Rizzi '09

    When: Thu, 1/21, 5-6pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Come to this informal panel session to learn about careers in computer science. Four alums will be there to talk about their career paths in CS, the job hunting process and how to prepare for it, and the transition from college to work. Bring your questions!

  • 2/4 Murat Dundar (1:20pm) -- Machine Learning in the Open World

    Machine Learning in the Open World
    Murat Dundar
    Purdue University

    When: Thu, 2/4, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Machine learning algorithms are now routinely used to build predictive models from data in wide range of applications. However, these algorithms are developed under highly controlled settings with a closed world assumption that do not reflect the dynamic nature of the open world surrounding us. For example object categories in the real world exhibits power law property; hence, in a randomly sampled data set collected for training purposes at a given time, no training examples are expected to be available for most of the object categories. A machine learning algorithm trained only with classes observed in the training set will misclassify all samples of unobserved classes into observed ones. This outcome creates a two-sided problem. First, the unknown class, which could potentially represent a significant abnormality such as a residual population of cancer cells in bone marrow or an emerging pathogenic bacteria strain present in food products, cannot be appropriately detected leading to potentially catastrophic consequences. Second, even if the unknown classes do not have any significance, misclassifying irrelevant samples into classes of practical importance raises doubts about the overall stability of the machine-learning systems, and will make these models highly vulnerable to adversarial attacks as has recently been the case with some well-established deep learning models. This talk will discuss machine learning algorithms that produce self-adjusting models that can accommodate new classes observed in data in offline as well as online learning scenarios. We categorize unobserved classes into two as known unknowns and unknown unknowns and discuss open-world machine learning in the context of zero-shot learning and open-set classification.

    Bio:
    Murat Dundar is an Associate Professor of Computer and Information Science at Indiana University Purdue University – Indianapolis. He received his BS degree from Bogazici University, Istanbul, Turkey, in 1997 and MS and PhD degrees from Purdue University, West Lafayette, IN, USA, in 1999 and 2003 respectively, all in Electrical Engineering. Between 2003 and 2008 He was with the CAD and Knowledge Solutions group of Siemens Health. At Siemens Health, he was involved in the development of a broad spectrum of computer aided diagnosis/detection applications including FDA-approved Lung and Colon CAD products. He has joined IUPUI Computer and Information Science Department as a tenure-track assistant professor in 2008, where he became an associate professor in 2014. His area of expertise is in machine learning with a special focus on open-world machine learning, where the goal is to replace the traditional brute-force approach of fitting a fixed model onto the data with more flexible models that can account for the non-stationary nature of real-world machine learning problems. His research is mainly driven by real-world problems in computer aided diagnosis/detection, hyper-spectral data analysis and remote sensing, and information technology. He was the recipient of the Data Mining Practice Price by ACM SIGKDD in 2009, Best Paper Award by International Association of Pattern Recognition in 2010, and Early Faculty CAREER Award by NSF in 2013. His research has been funded by NSF, NIH, and NASA.

  • 2/18 Ryan Gallagher (1:20pm) -- The Network Structure of Online Amplification

    The Network Structure of Online Amplification
    Ryan Gallagher
    Northeastern University

    When: Thu, 2/18, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Social media relies on amplification. It is at the heart of how marginalized communities voice injustices, how elected officials communicate public health guidance, and how misinformation proliferates through vulnerable populations. Each instance of amplification is a networked process emerging from many separate interpersonal interactions around an event, news story, or hashtag. Using the case study of #MeToo, I demonstrate how interactions between those who disclosed early in the hashtag campaign likely reduced the stigma of disclosure, allowing for further amplification of the hashtag. As it continued to be used, these disclosures transcended any single disclosure and coalesced into a larger network, composed of the core participants whose experiences of sexual violence were amplified by a larger periphery of bystanders and other survivors. I argue that this core-periphery structure is a fundamental signature of online amplification, and propose statistical models for how it can be identified empirically in networks. Finally, by applying these models back to the #MeToo case study, I demonstrate their effect on our ability to measure the reach of a hashtag activism event, highlighting the importance of accounting for the core-periphery network structure of amplification.

    Bio:
    Ryan Gallagher is a network science PhD candidate at Northeastern University. As a member of the Communication Media and Marginalization (CoMM) Lab at Northeastern's Network Science Institute, he studies how individuals use online communication networks to amplify their voices, and how that amplification resonates through online media ecologies. To do so, his research makes advances in network science and text-as-data methodology to develop new approaches for measuring the complexities of polarization, misinformation, and the networked public sphere. Ryan interned with Facebook Core Data Science and their Political Organizations & Society team, where he developed methods for identifying inauthentic coordinated information operations, and spent two summers as a visiting research assistant at the University of Southern California's Information Sciences Institute. He holds an MS in mathematics from the University of Vermont, where he worked with the Computational Story Lab at the Vermont Complex Systems Center, and a BA in mathematics from the University of Connecticut.

  • 2/25 Colleen Lewis (1:20pm) -- Recognizing and Responding to Bias and Microaggressions

    Recognizing and Responding to Bias and Microaggressions
    Colleen Lewis
    University of Illinois Urbana-Champaign

    When: Thu, 10/22, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Have you ever frozen - not knowing what to say - when you heard a comment or question about diversity in CS? We will play a research-based game to practice recognizing and responding to bias. The game invites players to respond to challenging scenarios related to subtle and not-so-subtle bias. For example, what might you say if your colleague said, "Women just don't like CS" or "There are so few Black and Latinx students in CS, it is a lot easier for them to get CS jobs." We're all responsible for learning to recognize and respond to bias -- and the game can provide opportunities to practice! We'll use breakout rooms to have players discuss how they would respond and you can access a copy of the game at www.csteachingtips.org/cards.

    Bio:
    Colleen Lewis is an Assistant Professor of computer science (CS) at the University of Illinois Urbana-Champaign. Lewis was previously the McGregor-Girand Associate Professor of CS at Harvey Mudd College. At the University of California, Berkeley, Lewis completed a PhD in science and mathematics education, an MS in computer science, and a BS in electrical engineering and computer science. Her research seeks to identify and remove barriers to CS learning and understand and optimize CS learning. Lewis curates CSTeachingTips.org, a NSF-sponsored project for disseminating effective CS teaching practices. Lewis has received the NCWIT.org Undergraduate Mentoring Award and the AnitaB.org Emerging Leader Award for her efforts to broaden participation in computing.

  • 3/8 ACM-W Seminar: Kerstin Dautenhahn (5pm) -- Social Robotics A New Generation of Robots Built for People

    Social Robotics A New Generation of Robots Built for People
    Kerstin Dautenhahn
    University of Waterloo

    When: Mon, 3/8, 5pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Compared to how people interact with computer and other electronic devices, something 'special' happens when humans come in contact and interact with physical robots. Human-Robot Interaction is a growing area of research where researchers try to understand how to design robotic systems that can interact with people. The talk will discuss the recent development of companion robots that can provide useful assistance to users in a socially acceptable manner. Such research focusses on the one hand on fundamental issues of human-robot interaction, learning and adaptation, but on the other hand is deeply inspired by concrete application areas and their requirements. The first part of the talk will introduce concepts and methodologies of developing social robots and discuss challenges.

    In the second part of the talk I will illustrate social robotics research in a particular application area that the speaker has studied since 2004, as part of the European projects Cogniron, LIREC and ACCOMPANY. Here the goal is to develop social robots as companions who can provide useful assistance (cognitive, physical and social) to people living in their own homes. One group of target users are elderly people who might benefit from a robot in their home in order to help them live independently. Such assistance is hoped to delay the move into a special care home, which is often associated with a loss of autonomy and a decline in wellbeing. Developing home assistance robots is a topic that is currently studied worldwide, and the talk will give examples from research in our own group. Our human-robot interaction research is primarily conducted in the University of Hertfordshire Robot House where we use the narrative framing technique in order to immerse study participants in our experimental scenarios. Such an approach bridges the gap between experimental laboratory environments and real homes, allowing for controlled studies in an ecologically valid environment. The talk will illustrate the scientific challenges and touch upon ethical and societal issues of such robot companion technology development.

    Bio:
    Since August 2018 Kerstin Dautenhahn has been Canada 150 Research Chair in Intelligent Robotics at University of Waterloo in Ontario, Canada. She has a joint appointment with the Departments of Electrical and Computer Engineering and Systems Design Engineering and is cross-appointed with the David R. Cheriton School of Computer Science at University of Waterloo. She is Visiting Professor at the University of Hertfordshire, UK. In Waterloo she is director of Social and Intelligent Robotics Research Laboratory (SIRRL). The main areas of her research are Human-Robot Interaction, Social Robotics, Assistive Technology and Artificial Life. She is Editor in Chief (jointly with Prof. Angelo Cangelosi - University of Manchester, UK) of the Journal Interaction Studies-Social Behaviour and Communication in Biological and Artificial Systems published by John Benjamins Publishing Company, Editorial Board Member of Adaptive Behavior, Sage Publications, Associate Editor of the International Journal of Social Robotics, published by Springer, and Associate Editor of IEEE Transactions on Cognitive and Developmental Systems (previously IEEE Transactions on Autonomous Mental Development). She is an Editor of the book series Advances in Interaction Studies, published by John Benjamins Publishing Company. Prof. Dautenhahn is on the Advisory Board of the journal AI and Society (Springer). She is a IEEE Fellow, member of ACM, and a Lifelong Fellow of AISB, as well as a member of the Executive Board of the International Foundation for Responsible Robotics. Since 2006 she has been part of the Standing Steering Committee of the IEEE conference RO-MAN (Human and Robot Interactive Communication).

  • 3/11 Ibrahim Abdul Rahmin (postponed to the fall)

    This seminar presentation had been postponed to fall 2021.

    Infrastructure Engineering: Large Scale, Fast Search
    Ibrahim Abdul Rahmin
    Salesforce

    When: Thu, 3/11, 1:20-2:15pm EST
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Infrastructure Engineering is the behind the scenes programming that makes internet scale companies like Microsoft, Google, Facebook able to serve billions of customers. This talk discusses Infrastructure Engineering for search: How do we take search from a single document, to many documents, to so many documents you need thousands of machines to host it all, and how do we manage search at that scale reliably?

    Bio:
    Ibrahim Abdul Rahim, is an Engineering Manager for Search at Salesforce. Under his leadership, his team builds systems that make search scalable, fast and relevant. We build upon Solr, an open source search solution. Last year the team shipped a search stack capable of indexing at a rate of 11M documents per minute. He was previously at Microsoft doing consumer search at Bing. He is an avid library book reader.

Fall 2020

  • 9/24 Enes Bilgin (1:20pm) -- Real-World Reinforcement Learning: Challenges and Opportunities

    Real-World Reinforcement Learning: Challenges and Opportunities
    Enes Bilgin
    Microsoft

    When: Thu, 9/24, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Reinforcement Learning (RL) is becoming increasingly popular among the machine learning community, thanks to the super-human level performance demonstrated in games like Atari 57, Go, Dota 2, and StarCraft. Many also see RL as a path to Artificial General Intelligence as it promises to tackle complex sequential decision-making problems under uncertainty, similar to animal learning and behavior. On the other hand, we are yet to see the hockey-stick growth in the value RL has the potential to deliver for real-world applications outside of game settings, such as in robotics, manufacturing, supply chain, and more. This talk focuses on what the obstacles in applying RL to real-world problems are, how the recent developments in literature are proposing solutions to these challenges, and how the big tech companies are incorporating RL into their AI-as-a-service offerings for their customers in the industry.

    Bio:
    Dr. Enes Bilgin is a Sr. AI Engineer in Microsoft’s Autonomous Systems organization, with expertise in Reinforcement Learning (RL). His work focuses on democratizing RL for a broad array of industries as part of Microsoft’s Project Bonsai. To this end, he develops “Machine Teaching” methods to enable subject matter experts to transfer their know-how to AI models in an intuitive and effective manner. Prior to Microsoft, Dr. Bilgin worked at MathWorks, AMD, and Amazon as a researcher and engineer. He was also an adjunct faculty at Texas State University and at the McCombs School of Business at the University of Texas at Austin. Dr. Bilgin holds a Ph.D. in Systems Engineering from Boston University.

  • 10/8 Fatih Camci (1:20pm) -- Prognostics and Health Management for Computer Scientists

    Prognostics and Health Management for Computer Scientists
    Fatih Camci
    Amazon

    When: Thu, 10/8, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Prognostics and Health Management (PHM) aims to increase safety and availability of engineering systems and reduce their ownership costs by predicting failures and avoiding unnecessary maintenance. PHM has been attracting industry and researchers from various fields in recent decade with its potential benefits. This presentation introduces the PHM field to computer science students and academics. The presentation will start with general introduction to the field with potential benefits. Then the challenges of PHM will be discussed as problems within computer science field.

    Bio:
    Dr. Fatih Camci works at Amazon Prime-Air as senior research scientist. He has worked in academia and industry since he received his PhD in 2005 from Industrial Engineering department in Wayne State University. He has MSc and BSc degrees in computer engineering. His research includes development of machine learning algorithms for failure forecasting in electro-mechanical systems. Today’s industry deals with many high-values assets that play critical role in human safety such as airplanes, helicopters, drones, nuclear power plants etc. Maintenance planning based using failure prediction algorithms plays a critical role in safety, reliability and ownership cost of these assets.

  • 10/15 Darren Strash (1:20pm) -- Engineering Fast Graph Algorithms with Data Reduction Rules

    Engineering Fast Graph Algorithms with Data Reduction Rules
    Darren Strash
    Hamilton College

    When: Thu, 10/15, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    Algorithm Engineering is a powerful methodology that pays homage to all aspects of algorithms, giving careful treatment to the interplay between theory, practice, and applications. An emerging trend in Algorithm Engineering is the discovery and application of data reduction rules, which can accelerate algorithms by up to multiple orders of magnitude (e.g., 100-10,000 times faster) by transforming an instance of a problem to a smaller equivalent instance. In this talk, I will introduce these critical tools in the context of my recent research in solving classical graph problems using data reduction rules. I will introduce the concept itself, simple data reductions, and their significant impact on several canonical graph problems. These include NP-hard problems (maximum independent set, maximum cut), as well as "easy" problems (minimum cut).

    Bio:
    Darren Strash is an assistant professor of computer science at Hamilton College, specializing in algorithms for large graphs. Before arriving at Hamilton, Strash worked at Intel, performed postdoctoral research at KIT in Germany, and taught at Colgate University. He and colleagues most recently won the Parameterized Algorithms and Computational Experiments (PACE 2019) challenge, a programming and research competition with participation from teams around the globe. He has co-authored over 40 peer-reviewed publications and received two best paper awards.

  • 10/22 Nick Seaver (1:20pm) -- Algorithms and/as Culture

    Algorithms and/as Culture
    Nick Seaver
    Tufts University

    When: Thu, 10/22, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:
    The algorithms that shape culture, filtering what we see and hear, are often understood as inhuman forces. This talk describes how these systems are full of people who impart their own points of view on the software they build. Drawing on anthropological fieldwork with the developers of music recommender systems, it explores the consequences of thinking about algorithms in broader cultural contexts and as cultural objects in their own right.

    Bio:
    Nick Seaver is an assistant professor in the Department of Anthropology and the Program on Science, Technology, and Society at Tufts University in Medford, MA. He studies how technologists make sense of cultural concerns such as taste and attention.

  • 11/5 Dan Sheldon (1:20pm) -- Bayesian Forecasting of COVID-19

    Bayesian Forecasting of COVID-19
    Dan Sheldon
    University of Massachusetts Amherst

    When: Thu, 11/5, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)

    Abstract:

    COVID-19 has infected more than 41 million people worldwide, with more cases in the US than any other country. COVID-19 models can help provide “situational awareness” of disease prevalence and make short-term forecasts to provide actionable information to public health planners.

    In early March 2019, I was on sabbatical catching up on research projects. In late March, I connected with my friend Nick Reich, who leads a team of infectious disease forecasters at UMass. Since then I have led a small team developing the “MechBayes” (Mechanistic Bayesian) COVID-19 forecast model using tools from computer science and statistics. We submit weekly forecasts to the COVID-19 Forecast Hub, which are then provided to the CDC and used in an ensemble forecast. Our model is one of several featured on the FiveThirtyEight website.

    In this talk, I will provide an overview of how we combined classical epidemiology models with modern probabilistic programming to create these forecasts. I will also describe the broader efforts of the COVID-19 Forecast Hub to standardize, collect, and aggregate forecasts by building ensemble models.

    Bio:
    Dan Sheldon is an associate professor in the College of Information and Computer Sciences at the University of Massachusetts Amherst and the Department of Computer Science at Mount Holyoke College. His research investigates fundamental problems in machine learning and applied algorithms motived by large-scale environmental data, dynamic ecological processes, and real-world network phenomena (and now epidemiology).

  • 11/12 CSC 498 Poster Session (1:20pm)

    CSC 498 Poster Session
    CSC 498 students
    Union College

    When: Thu, 11/12, 1:20-2:15pm EDT
    Where: the Zoom link will be sent out by email
    (If you would like to be added to the distribution list, please contact Kristina Striegnitz at striegnk@union.edu)