All Seminars

• Upcoming Seminars
• All Seminars

Title: Knowledge Discovery of Graph Transformation Patterns by Deep Generative Models and Optimization
Seminar: Computer Science
Speaker: Dr. Liang Zhao, George Mason University
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-02-12 at 11:30AM
Venue: Atwood 215
Download Flyer  Add to Calendar
Abstract: Inspired by the tremendous success of deep generative neural networks on modeling and generating continuous data like image and audio, in recent couple of years, deep graph generative learning is becoming a promising domain which focuses on generating graph-structured data. Most of them are unconditioned generative models which has no control on modes of the graphs being generated. Going beyond that, in this presentation, I will talk about a new topic named Deep Graph Transformation: given a source graph, we want to infer a target graph based on their underlying global and local transformation mapping. By automatically interpreting such generative transformation process, we aim to distill new rules and patterns of the underlying transformation mechanism. Deep graph transformation could be highly desirable in many promising applications on network synthesis and prediction, such as chemical reaction simulation, brain network modeling, and protein structure prediction. I will introduce our recent progress on new structured learning frameworks, convolution and deconvolution operations, and intepretability enhancement techniques for deep graph transformation. Furthermore, I will further talk about the computational bottlenecks of current optimization techniques for training deep neural networks, especially for complex, large data such as graphs. I will introduce our recent work on gradient-free optimization techniques for deep neural network optimization based on deep alternating optimization, which aims to overcome several existing fundamental drawbacks such as gradient vanishing, low concurrency, and biological implausibility. Bio: Dr. Liang Zhao is an assistant professor at the Department of Information Science and Technology at George Mason University. He obtained his PhD degree in 2016 from Computer Science Department at Virginia Tech in the United States. His research interests include data mining, artificial intelligence, and machine learning, with special interests in spatiotemporal and network data mining, deep learning on graphs, nonconvex optimization, and interpretable machine learning.

Title: Towards Safe Machine Learning for Health
Seminar: Computer Science
Speaker: Shalmali Joshi, PhD, Vector Institute
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-02-05 at 11:30AM
Venue: Atwood Chemistry Center 215
Download Flyer  Add to Calendar
Abstract: Advances in machine learning (ML) have led to modest success in clinical healthcare, as opposed to fields like computer vision. Many research challenges remain that are critical for safe deployment of ML methods for clinical decision support. One of these challenges is related to the ability of ML models to learn clinically relevant and actionable insights. In the first part of my talk I focus on contextualizing these concerns in terms of explainability of ML methods and the widening discrepancy between current ML explainability research and clinician expectations. One such severely understudied problem is determining the individualized feature relevance in time series models. We propose a method to quantify feature importance in this setting by leveraging deep generative models and demonstrate its efficacy on simulated and real world data. While disparities in healthcare have been well documented, the effect of using biased data in ML, especially for causal effect estimation, needs more scrutiny. To that effect, in the second part of my talk, I will describe a method we propose to evaluate the reliability of state of the art ML based counterfactual regression models in the presence of treatment and outcome disparity and relating their efficacy to underlying data generative settings and awareness of the source of disparity. We end with a vision of a research goals toward addressing safe and fair deployment of ML for health.

Title: Rethinking Text Generation Models and How to Train Them
Seminar: Computer Science
Speaker: Sam Wiseman, PhD, Toyota Technological Institute at Chicago
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-02-04 at 10:30AM
Venue: Mathematics and Science Center: W502
Download Flyer  Add to Calendar
Abstract: Much of the recent empirical success in natural language processing has relied on the use of neural networks to compute expressive, global representations of sentences and documents. However, incorporating such representations into systems that produce outputs with combinatorial structure, such as text, may require rethinking both our models and how we train them. In terms of training, I will argue for training text generation models to search for optimal outputs, which will address some of the shortcomings of standard maximum likelihood-based training. In terms of modeling, I will argue that standard text generation models are difficult to interpret and to control, and I will suggest a model that automatically induces discrete template-like objects, which can be used for controlling and interpreting generation.

Title: Enabling Mixed Reality for Low Vision
Seminar: Computer Science
Speaker: Yuhang Zhao, Cornell University
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-02-03 at 11:30AM
Venue: W507
Download Flyer  Add to Calendar
Abstract: Visually impaired people are marginalized by inaccessible social infrastructure and technology, facing severe challenges in all aspects of their life. Mixed reality (MR) technology has the capability of incorporating virtual information into the physical environment, presenting a unique opportunity to augment the world for visually impaired people. However, it also creates a virtual world that is currently vision dominant, which can cause more accessibility issues. I strive to explore how MR technology can empower visually impaired people, providing them equal access to both the real and virtual worlds. In this talk, I will focus on people with low vision, who have visual impairments but are not blind. I will discuss how I leverage MR technology to address both the real-world challenges and the virtual world accessibility for low vision. To solve the real-world challenges that low vision people face, I design and build intelligent MR systems to directly enhance their visual ability by providing visual augmentations. For example, I built a head-mounted MR system that presented visual cues to orient users’ attention in a visual search task, as well as a MR system on HoloLens that generated visualizations to support safe stair navigation. Meanwhile, to foster the accessibility of the virtual world generated by MR, I adapted the real-world low vision aids and technology to the virtual world, creating a set of tools that enhance virtual reality (VR) applications for low vision people. To universally apply these tools, I developed both a plugin to modify an existing VR application post hoc, and a Unity toolkit that enables developers to build more accessible VR applications. I will conclude my talk by highlighting my future research directions, such as building MR systems for multi-user scenarios (e.g., social interaction) and diverse disabilities (e.g., autism), and constructing general MR accessibility frameworks. Bio: Yuhang Zhao is a sixth-year PhD candidate in Information Science at Cornell University. Her research interests lie in human-computer interaction (HCI), accessibility, and augmented and virtual reality. She designs and builds intelligent interactive systems to enhance human abilities. She has published at many top-tier conferences and journals in the field of HCI (e.g., CHI, UIST, ASSETS), and has received 3 U.S. and international patents. She has interned at Facebook, Microsoft Research, and Microsoft Research Asia. Her work received two best paper honorable mention awards at the SIGACCESS Conference on Computers and Accessibility (ASSETS) and has been covered by various media outlets (e.g., TNW, New Scientist). She received her B.A. degree and M.S. degree with distinction on thesis in Computer Science at Tsinghua University.

Title: Human Compatible: Making Decisions From Human Knowledge and Preferences
Seminar: Computer Science
Speaker: Rupert Freeman, PhD, Microsoft Research, NYC
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-01-31 at 10:30AM
Venue: MSC W201
Download Flyer  Add to Calendar
Abstract: Even in the age of big data and machine learning, human knowledge and preferences still play a large part in decision making. For some tasks, such as predicting complex events like recessions or global conflicts, human input remains a crucial component, either in a standalone capacity or as a complement to algorithms and statistical models. In other cases, a decision maker is tasked with utilizing human preferences to, for example, make a popular decision over an unpopular one. However, while often useful, eliciting data from humans poses significant challenges. First, humans are strategic, and may misrepresent their private information if doing so can benefit them. Second, when decisions affect humans, we often want outcomes to be fair, not systematically favoring one individual or group over another. In this talk, I discuss two settings that exemplify these considerations. First, I consider the participatory budgeting problem in which a shared budget must be divided among competing public projects. Building on classic literature in economics, I present a class of truthful mechanisms and exhibit a tradeoff between fairness and economic efficiency within this class. Second, I examine the classic online learning problem of learning with expert advice in a setting where experts are strategic and act to maximize their influence on the learner. I present algorithms that incentivize truthful reporting from experts while achieving optimal regret bounds. Bio: Rupert Freeman is a postdoc at Microsoft Research New York City. Previously, he received his Ph.D. from Duke University under the supervision of Vincent Conitzer. His research focuses on the intersection of artificial intelligence and economics, particularly in topics such as resource allocation, voting, and information elicitation. He is the recipient of a Facebook Ph.D. Fellowship and a Duke Computer Science outstanding dissertation award.

Title: Multi-Facet Contextualized Graph Mining with Cube Networks
Seminar: Computer Science
Speaker: Carl Yang, PhD, University of Illinois, Urbana Champaign
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-01-29 at 11:30AM
Venue: MSC W507
Download Flyer  Add to Calendar
Abstract: Graph data are ubiquitous and indispensable in a variety of high-impact data mining problems and applications, due to its natural and unique modeling of interconnected objects. However, real-world graph data are often massive, complex, and noisy, challenging the design of both effective and efficient knowledge discovery frameworks. In this talk, I will present our recent progress on multi-facet contextualized graph mining, centered around the objective of multi-modal data integration across different domains. In particular, I will focus on (1) a new data model of cube networks, which organizes massive complex networks into controllable small subnetworks with clear structures and semantics under multi-facet contexts; (2) a few algorithmic examples on what can be done on top of cube networks. Beyond that, I will also briefly give examples on how to construct cube networks from existing data models like attributed heterogeneous networks, and what real-world impact cube networks can make on industry-level applications. Finally, I will conclude with some visions and future plans regarding learning with cube networks. Bio Carl Yang is a final-year Ph.D. student with Jiawei Han in Computer Science at University of Illinois, Urbana Champaign. Before that, he received his B.Eng. in Computer Science at Zhejiang University under Xiaofei He in 2014. In his research, he develops data-driven, weakly supervised, and scalable techniques for knowledge discovery from massive, complex and noisy network (graph) data. His interests span graph data mining, network data science, and applied machine learning, with a focus on designing novel graph analysis and deep learning frameworks for the construction, modeling, and application of real-world network data, towards tasks like conditional structure generation, contextualized network embedding, graph-aided recommendations, and so on. Carl’s leading-author research results have been published and well-cited in top conferences like KDD, WWW, NeurIPS, ICDE, WSDM, ICDM, CIKM, ECML-PKDD, SDM and ICML.

Title: Risk-Sensitive Safety Analysis and Control for Trustworthy Autonomy
Seminar: Computer Science
Speaker: Margaret P. Chapman, University of California, Berkeley
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-01-28 at 10:00AM
Venue: MSC W507
Download Flyer  Add to Calendar
Abstract: Methods for managing dynamic systems typically invoke one of two perspectives. In the worst-case perspective, the system is assumed to behave in the worst possible way; this perspective is used to provide formal safety guarantees. In the risk-neutral perspective, the system is assumed to behave as expected; this perspective is invoked in reinforcement learning and stochastic optimal control. While the worst-case perspective is useful for safety analysis, it can lead to unnecessarily conservative decisions, especially in settings where uncertainties are non-adversarial. The risk-neutral perspective is less conservative and useful for optimizing the system’s performance on average. However, optimizing average performance is not guaranteed to protect against harmful outcomes and thus is not appropriate for safety-critical applications. In this talk, I will first present an analytical and data-driven computational toolkit for managing triple-negative breast cancer that I have developed with cancer biologists at Oregon Health and Science University by invoking the worst-case perspective. In addition to providing biological insights about breast cancer and theoretical insights about switched systems, this work has motivated the need for new mathematical methods that facilitate less conservative but still protective control of dynamic systems. Towards this aim, I have devised a risk-sensitive mathematical method for safety analysis that blends the worst-case and risk-neutral perspectives by leveraging the Conditional Value-at-Risk measure. The next part of my talk will focus on this new risk-sensitive mathematical method and its application to evaluating safety of urban water infrastructure in joint work with water resources specialists at the Berkeley Water Center and Tufts University. In the last part of my talk, I will present the aims of my future research program “Decision Analysis for Trustworthy Autonomy” that will extend the theories of control, risk, and learning to address safety-critical challenges in society. Bio: Margaret Chapman is a PhD Candidate advised by Dr. Claire Tomlin in Electrical Engineering and Computer Sciences (EECS) at UC Berkeley. She is grateful to be a recipient of the Fulbright Scholarship, the NSF Graduate Research Fellowship, and the Berkeley Fellowship for Graduate Study. She earned her BS degree (with Distinction) and her MS degree in Mechanical Engineering from Stanford University. Margaret’s research interests are the development of data-driven dynamic models and mathematical decision analysis methods for safety-critical stochastic systems with applications to sustainable cities and healthcare. Margaret is delighted to be a participant of 2019 Rising Stars in EECS, and she aims to become a professor at a university with strong interdisciplinary research. "https://www.margaretpfeifferchapman.com

Title: Just, Equitable, and Efficient Algorithmic Allocation of Scarce Societal > Resources
Seminar: Computer Science
Speaker: Sanmay Das, Washington University in St. Louis
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-01-27 at 11:30AM
Venue: MSC W507
Download Flyer  Add to Calendar
Abstract: Demand for resources that are collectively controlled or regulated by society, like social services or organs for transplantation, typically far outstrips supply. How should these scarce resources be allocated? Any approach to this question requires insights from computer science, economics, and beyond; we must define objectives (foregrounding equity and distributive justice in addition to efficiency), predict outcomes (taking causal considerations into account), and optimize allocations, while carefully considering agent preferences and incentives. In this talk, I will discuss our work on weighted matching and assignment in two domains, namely living donor kidney transplantation and provision of services to homeless households. My focus will be on how effective prediction of the outcomes of matches has the potential to dramatically improve social welfare both by allowing for richer mechanisms and by improving allocations.

Title: Relational Learning Methods and Complex Networks with Applications to Health Informatics
Seminar: Computer Science
Speaker: Pablo Robles-Granda, PhD, University of Notre Dame
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2020-01-24 at 10:30AM
Venue: MSC W201
Download Flyer  Add to Calendar
Abstract: Statistical relational models are widely used in network science to reason about the properties of complex systems — where the nodes represent entities (e.g., individuals, disease vectors, patients etc.) and the links represent relationships (e.g., friendship, contagion paths, bio- and social-markers, etc.). A type of relational models is generative network models that aim to capture the observed characteristics of real world networks to reproduce them via sampling. To acquire a better understanding of the underlying properties of the system (e.g., a social network), many challenges must be overcome. An example of these challenges is that a good descriptive model of networks not always facilitate sampling. In order to overcome this challenge, it is important to develop relational models that facilitate sampling methods for networks with and without vertex-attributes. However, this task remains a challenging problem because most current methods work with relatively simple or specific models. Another challenge is to understand the dynamics that appear in a complex system and how these dynamics evolve and what are the dependencies across components of the system that drive these dynamics. In my talk, I will discuss some developments for these two types of problems starting with an application to health informatics. First, nowadays, the popularity of machine learning has produced a significant body of research on social and medical sciences. In my work, I have developed mathematical descriptions of health and wellness of the individuals. I investigated various techniques of network science to extract the graph-theoretic bio-markers. Also, I applied machine learning to predict stress and wellness variables. With my colleagues, we showed how to use these additional biomarkers to improve the accuracy of prediction of wellness of individuals. I will present a few other examples where I apply some methodological insights acquired from relational learning. Second, I will present observations about a number of recent models that share a common structure. This structure allows to increase both the variance of the models and the space of characteristics possible to be modeled. Building on these results I will discuss general strategies to combine models and density functions of the node attributes to sample attributed-networks. Since in this structure the probability mass is allocated to certain regions of the network space it is hard to identify candidate networks to sample attributed networks. I will present constrained sampling to bias the search of edge-candidates using the topological space with higher likelihood. Finally, identifying the laws that govern a complex system based on its network representation is an open problem. In my work, I identified strategies to represent changes in a dynamical system using its footprint data. The persistent challenge is that most approaches for incident detection in complex non-linear systems use nominal systems that rely on domain expertise. With my colleagues, we have studied the problem of incident detection in non-linear systems using the trajectory of the causal relationships described only by the time-series data.

Title: Intelligent High-Performance Networks via INCA
Seminar: Computer Science
Speaker: Ryan Grant,
Contact: Dorian Arnold, dorian.arnold@emory.edu
Date: 2020-01-24 at 1:30PM
Venue: MSC E300 Planetarium
Download Flyer  Add to Calendar
Abstract: In this talk we will describe a deadline-free general compute model for network endpoints called INCA: In-Network Compute Assistance. INCA builds upon contemporary NIC offload capabilities to provide on-NIC, deadline-free, general-purpose compute capacities that can be utilized when the network is inactive. We provide a detailed design for extending existing hardware to support this model. We will demonstrate where INCA fits in the existing smart network ecosystem and detail why it is a major departure from existing approaches. We will describe the new challenges that we overcame to invent INCA, detailing why creating a deadline-free general compute model for networks for HPC has not previously been explored. We will offer use cases for areas where INCA can be used and expanded, with special emphasis on using INCA to solve problems in the sciences and enabling new approaches to distributed machine learning that will be possible with INCA.